SERIES STANDARDS FOR INFANT RESPIRATORY FUNCTION TESTING ERSATSTASK FORCEEdited by J Stocks and J GerritsenNumber 4 in this Series

Tidal breath analysis for infant pulmonary function testing

JHT Bates G Schmalisch D Filbrun J Stocks+ on behalf of theERSATS Task Force on Standards for Infant Respiratory Function Testing

Tidal breath analysis for infant pulmonary function testing JHT Bates G SchmalischDFilbrun J Stocks on behalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing ERS Journals Ltd 2000ABSTRACT The aim of this position paper is to provide recommendationspertaining to software and equipment requirements when analysing tidal breathingmeasurements in infants These guidelines cover numerous aspects includingterminology and definitions equipment data acquisition and analysis and reportingof results and highlight areas in which further research is needed before consensuscan be reached

When collecting tidal breathing data in infants and children equipment dead spaceand resistancemust be minimized all sources of leak eliminated and a flowmeter withappropriate frequency response and linearity employed Inspired gases should becorrected to body temperature barometric pressure and saturated with water vapourconditions and efforts made to eliminate the various sources of drift in volume thatcan occur In addition the analogue-to-digital converter used to sample data must becapable of adequately resolving the highest and lowest flows required by the study Anadequate sampling rate must be used 50ndash100 Hz may be sufficient for thedetermination of timing and volume parameters especially in older infants but ratesof 200 Hz are recommended for analysis of the tidal breathing flowvolume loop andother sensitive parameters such as time to peak tidal expiratory flowexpiratory time

The potentially most troublesome aspect of tidal breath analysis from thecomputational point of view is the identification of the beginning and end ofinspiration and expiration

Once methods and equipment for the measurement and analysis of tidal breathingin infants have been standardized there is an urgent need to establish appropriatereference ranges for various key parameters so that they may be used more effectivelyin the clinical setting Implementation of these recommendations should help to ensurethat such measurements are as accurate as possible and that more meaningfulcomparisons can be made between data collected in different centres or with differentequipmentEur Respir J 2000 16 1180ndash1192

Vermont Lung Center the University ofVermont Colchester VT USA Dept ofNeonatology (Charite) Humboldt Uni-versity Berlin Germany Infant Pul-monary Laboratory Childrenrsquos HospitalColumbus OH USA +Portex Anaesthe-sia Intensive Therapy and RespiratoryMedicine Unit Institute of Child HealthLondon UK

Correspondence JHT Bates208 South Park DriveSuite 2Colchester VT 05446USAFax 1 8026568932

Keywords Infantsrespiratory function testsstandardizationtidal breathing

Received March 10 2000Accepted after revision June 13 2000

This work was supported by a grant fromthe European Respiratory Society and bydonations from Glaxo Wellcome (UK) andGlaxo Wellcome AB (Sweden)

The present document represents one of a series [1ndash6]that is being produced by the European RespiratorySocietyAmerican Thoracic Society Task Force on stan-dards for infant respiratory function testing The aim ofthis task force is to summarize what is currently seen to begood laboratory practice and to provide recommenda-tions for both users and manufacturers of infant lungfunction equipment and software These recommenda-tions have been developed after widespread communica-tion on an international level and are directed towards

future developments in this field including the use ofmore automated and standardized equipment than hasbeen used in the past It is recognized that this documentwill need to be updated regularly in response to advancesin both technology and knowledge regarding the appli-cation and interpretation of these tests under differentcircumstances In the meantime every attempt has beenmade to avoid being too prescriptive in order to allow forfuture developments while at the same time offeringguidance as to minimum standards for those developing

Previous articles in this series No 1 U Frey J Stocks A Coates PD Sly J Bates on behalf of the ERSATS Task Force onStandards for Infant Respiratory Function Testing Specificationsfor equipmentused for infants pulmonary function testing Eur RespirJ 2000 16 729ndash738 No 2 PD Sly R Tepper M HenschenM Gappa J Stocks on behalf of the ERSATS Task Force on Standardsfor Infant Respiratory Function Testing Tidal Forced ExpirationsEur Respir J 2000 16 739ndash746 No 3 U Frey J Stocks P Sly JBates on behalf of the ERSATS Task Force on Standards for Infant Respiratory Function Testing Specifications for signal processingand data handling used for infant pulmonary function testing Eur Respir J 2000 16 739ndash746

Eur Respir J 2000 16 1180ndash1192Printed in UK ndash all rights reserved

Copyright ERS Journals Ltd 2000European Respiratory Journal

ISSN 0903-1936

equipment and performing tests It is anticipated thatacceptance and application of these recommendationswill be of particular value when attempting to comparedata between centres develop or use reference data orparticipate in multicentre trials which use tidal breathingparameters as outcome measures

An infantrsquos breathing pattern measured during tidalbreathing contains significant physiological informationpertaining to a number of processes related to respiratorycontrol and pulmonary mechanical function Such in-formation is encapsulated within a number of conceptually

straightforward parameters The most fundamental para-meters contained in the flow and volume signal tidalvolume (VT) respiratory frequency (fR) and inspiratory (tI)and expiratory time (tE) are shown in figure 1a and b fRwhich is calculated as 60(tI+tE) is also referred to asrespiratory rate (RR) These basic parameters can be usedto calculate quantities pertaining to the more detailedaspects of the pattern and magnitude of tidal breathingsuch as the time to peak tidal expiratory flow (tPTEF) as aproportion of tE (tPTEFtE) minute ventilation (VrsquoE) meaninspiratory flow (VTtI) and the duty cycle (tItotal breathtime (ttot)) In addition tidal breathing flowvolume loopsmay be plotted for visual inspection and evaluation ofparameters such as the volume expired up to the time ofpeak expiratory flow (VPTEF) or defined flow in relationto exhaled volume (fig 1c) It may also be useful toderive certain composite parameters such as VPTEFrelated to VT (VPTEFVT) However the purpose of thisdocument is not to provide a theoretical background tothe factors determining tidal breathing patterns duringearly life nor to comment on the potential clinicalsignificance (or irrelevance in some instances) of theinnumerable parameters that can be calculated fromrecordings of tidal breathing For this the reader isdirected to the existing literature much of which hasbeen summarized recently [7] The intent is rather tostress the various factors during data acquisition analysisand reporting of tidal breathing parameters that can leadto systematic errors between different recording systems

Obtaining tidal breathing parameters requires nothingmore than the measurement of flow or volume at the mouthand nose during a period of regular breathing Howeverthere are a number of considerations pertaining to theanalysis of tidal breathing data that make it somewhat lessthan trivial The most problematic issue from a practicalpoint of view is the precise detection of the onset ofinspiration and expiration The problem of automaticallysegmenting breaths into their inspiratory and expiratoryphases is thus considered in the present document Otherpractical issues such as the collection of airflow data andthe derivation of a drift-free volume signal from such dataare also considered Some of these practical issues areinfluenced by the task in hand For example investigationsinto the control of breathing require the determination ofnot only the magnitude of but also the variability in VT tIand tE On the one hand this requires measurement of alarge number of breathing cycles which cannot be per-formed with the conventional combination of face maskand pneumotachometer (PNT) due to their relatively highapparatus dead space thus alternative approaches areusually required [8] On the other hand a relatively lowsample rate (10ndash50 Hz) will often suffice for such studiesBy contrast when attempting to analyse the mechanicalcomponents of the tidal breathing signal a relatively lownumber of consecutive reproducible and artefact-freebreaths need to be recorded but at a higher sampling rate

Because of the complexities of using noninvasive bodysurface measurements to obtain quantitatively accurateassessments of tidal breathing parameters and the lack ofany standardized approach to such measurements thecurrent document is limited to direct measurements ofairflow and volume at the airway opening Neverthelesssome of the recommendations within may be pertinent ifalternative approaches are used

Volu

me

a)

tEtI

Flow

b)

tPTIF

PTIF

PTEF

tPTEF

Time

Insp Exp

ttot

VT

Insp Exp

c)

Flow

Volume

075 05 025 01VT

VPTEF

PTEF TEF50

TIF50PTIF

Insp

Exp

Fig 1 ndash Tidal breathing parameters of a) the flow and b) volumesignals and c) the flowvolume loop The tidal expiratory (TEF50) andinspiratory flow when 50 of tidal volume (VT) remains in the lung(TIF50) are shown other values of TEFx and TIFx correspond to thedifferent VT on the scale bar Insp inspiration exp expiration tIinspiratory time tE expiratory time ttot total breath time PTIF peaktidal inspiratory flow PTEF peak tidal expiratory flow tPTEF time topeak expiratory flow tPTIF time to peak inspiratory flow VPTEFvolume expired before PTEF attained

1181TIDAL BREATH ANALYSIS IN INFANTS

General procedures

In the present paper tidal breathing is taken to be thenatural physiological state of undisturbed regular breath-ing Measurements are commonly performed during quietsleep which in older infants is usually associated withactive inspiration and passive expiration In order to recordperiods of undisturbed breathing there should be neg-ligible influence of the measuring equipment [9] thustidal breathing should be assessed prior to more complexinvestigations For example although it is certainly pos-sible to measure tidal breathing parameters immediatelyafter insertion of an oesophageal catheter the valuesobtained will be significantly different from those mea-sured prior to catheter insertion Meaningful comparisonof tidal breathing parameters within and between infantsor interpretation with respect to reference values areonly possible if the measurement conditionsare the sameCareful documentation of measurement conditions andequipment characteristics during tidal breathing measure-ments is therefore essential

Contrary to common belief collection of undisturbedquiet tidal breathing data is not a trivial undertaking aseven the application of a face mask can stimulate breath-ing [10 11] Consequently measurements should onlybe made after a sufficient adaptation period has beenallowed to elapse after attachment of the measurementequipment This is not a problem if there is negligibleadditional dead space involved such as with the flow-through technique [8] However when using a conven-tional mask and PNT the apparatus dead space limits theduration of measurements [12] Furthermore the smallerthe infant the greater the impact of the apparatus deadspace Careful preparation is therefore necessary tominimize the adaptation period especially when studyingnewborns

Careful use of equipment in order to ensure patientsafety remains the responsibility of the operator [1]Routine safety measures in the pulmonary functiontesting laboratory include 1) availability of full resuscita-tion equipment including suction at the site of infantlung function testing plus a suitable alarm system 2)the presence of two individuals (other than parents)during testing one of whom has prime responsibility forthe infantrsquos wellbeing the infant must never be leftunattended 3) continuous physiological monitoring (andideally recording) including at least pulse oximetry 4)use of transparent face masks and 5) adherence to thehospital-specific protocol for sedation or anaesthesia

Further details regarding measurement conditionswhich may influence infant safety or the accuracy andreproducibility of results have been published previously[5 13] The measurement procedure can be summarizedas follows 1) The infant should be lying supine with theneck andor shoulders supported in the midline in slightextension This position should be stabilized by use of aneck roll or head ring If an alternative posture is usedthis should be stated clearly in any publication 2) Theinfant should be fed dry clean and comfortably clothed3) The face mask should cover mouth and nose and beplaced with minimal pressure while also being airtightTo prevent any air leaks a thin sealing ring of siliconputty is often helpful

Equipment

Airflow and volume determined by numerical integra-tion of airflow are the basic signals of tidal breathinganalysis (fig 1) Airflow is usually measured at the airwayopening using a PNT connected to a face mask Althoughboth respiratory inductance and face-out body plethys-mography have been used to record tidal breathingsignals the former only provides quantitatively accuratemeasurements if there has been previous calibration usinga PNT and the measurement conditions remain verystable [14] whereas the latter is too expensive andcumbersome for routine bedside application A detaileddescription of devices available for the measurement offlow and hence volume has been published recently[15] The reader is also referred to the document in thepresent series on specifications for equipment and soft-ware used for infant pulmonary function testing forfurther details and justification of the recommendationspresented below [1 3]

Face mask

Different types of mask are used based largely on localpreference and availability The dead space of the facemask and flowmeter may have a marked influence onthe pattern of breathing with respect to both magnitudeand timing thereby altering the very parameters that theinvestigator is trying to assess In preterm and sickneonates even when a small face mask and a very-low-dead space flowmeter are used the total apparatus deadspace may exceed the infantrsquos own dead space and pre-sent a considerable load thereby precluding anything butbrief intermittent recordings In such infants the use of thedead-space-free-flow-through technique is recommended[8]

In order to aid international standardization thefollowing recommendations are made 1) report brandand size of mask and whether the mask has an air-filledcuff since the mask dead space depends significantly onthe cuff pressure 2) document other aids to improvingairtight placement (eg silicon putty or Vaseline) 3) mini-mize dead space and report whether measures to reducemask volume are used (eg shortening of the mask port orputting silicon putty in the mask) and 4) measure the deadspace of the mask by water displacement and subtract 50of this volume to estimate the effective dead space [1 16]

Flowmeter

The conventional transducer employed for measurementof respiratory flow is the combination of PNT and diffe-rential pressure transducer Attention is restricted to thisdevice in the present document although several newdevices such as ultrasonic flowmeters and hot wireanemometers [17] are currently being developed andvalidated for use in infant respiratory function measure-ments A number of general considerations concerninginstrumentation and measurement technique pertain tothe measurement of flow and its recording using acomputer These are as follows 1) The flowmeter shouldbe a low-resistance low-dead-space device Flowmeters

1182 JHT BATES ET AL

for measurements in preterm infants should have a deadspace of lt15 mL Unfortunately the apparatus deadspace of most modern devices arises largely due to thenecessary connecting ports and so possibilities for deadspace reduction are limited In any case the connectionbetween PNT and mask must be minimized withoutdisturbing the linearity of the flowmeter [15] 2) Mini-mizing the resistance of the infant lung function equip-ment is important since the overall resistance of theequipment may not only dramatically change the respi-ratory pattern in spontaneously breathing babies butalso interfere with triggering devices in those who areventilated Any significant increase in resistance in-creases the expiratory time constant and potentiallyinfluences the end-expiratory level This in turn affectsany measurements that are volume-dependent includingvarious tidal breathing parameters The need to designfuture apparatus with as low a resistance as possiblewithin the constraints of simultaneously attaining a lowdead space and high resolution cannot be overempha-sized 3) The combined resistance of the apparatus (in-cluding any valves capnographs etc) should be lt20of the infantrsquos intrinsic resistance at the mean flowslikely to be encountered [1] Thus as a rough guide thecombined apparatus resistance should not exceed 12kPaL-1s at 50 mLs-1 in spontaneouslybreathing preterminfants 07 kPaL-1s at 100 mLs-1 in term neonates and05 kPaL-1s at 300 mLs-1 in infants and young children4) The response of the flowmeter should be linear overthe range of flows encountered The extent to which aPNT remains linear over an extended range is criticallydependent on design features such as whether it is acapillary- screen- or variable orifice-type device and thegeometry of any integral connections It is thereforeessential that the manufacturer provides accurate detailsand that the user checks the range of flows over whichthe flowmeter provides accurate recordings The approx-imate linearity ranges required for various sizes of infantare 0ndash100 mLs-1 in preterm infants and neonates of 2ndash4kg 0ndash300 mLs-1 in infants of 4ndash10 kg and 0ndash500 mLs-1

in preschool children of 11ndash15 kg In practice flowmeterswith a linear range of 0ndash10 Lmin-1 are commonly used inpreterm infants and neonates whereas those with a rangeof 0ndash35 Lmin-1 are used for obtaining tidal breathingmeasurements in older infants and young children 5) If aPNT has a nonlinear response over the desired flowrange it may be possible to effectively linearize it bycharacterizing the response and inverting it digitally Thiscan reduce dead space by allowing the use of a smaller-calibre PNT However it must be ensured that theresponse characteristics of the device remains constantover a prolonged period after repeated disinfection andon exposure to different respired gases 6) The flowmetershould have a flat frequency response over a frequencyrange sufficient to encompass the majority of the power inthe measured signals [1] For tidal breathing signals it isprobably sufficient to have a flat frequency response upto 10 Hz If the transducer itself does not have a flatfrequency response over this range it may be possibleto render it flat by digital compensation of the sampleddata [18] however this is only possible if the responseof the device is linear 7) If a PNT with metal screensor capillaries is used it should be heated to body tem-perature to avoid condensation on the resistive element

Major changes in screen resistance and hence measuredflows can occur within lt1 min of placing an unheatedPNT into a ventilator circuit thus this practice is stronglydiscouraged 8) Despite such heating PNTs with screensor capillaries in the ventilator circuit are highly suscep-tible to obstruction by secretions leading to falsely highmeasured flows and possible danger to the patientTherefore these PNTs should only be used by qualifiedpersonnel while the patient is under direct observation[19] 9) The geometry of the connectors on either side ofthe PNT screen affects the overall pressureflow charac-teristics of the device It is therefore important that theconnectors be as symmetrical as possible on either side ofthe PNT and that the PNT is calibrated in situ in exactlythe same configuration as that to be used with the subject[15] 10) If the inspired gas differs significantly fromroom air (eg by increased inspiratory oxygen fraction(FIO2)) it may be of different viscosity to room air andtherefore have different PNT calibration factors In such acase either the PNT should be calibrated with the inspiredgas or the room air calibration factors should be scaled bythe relative differences in gas viscosity [15] For measure-ments during artificial ventilation continuous FIO2correction at the bedside is advantageous [19] Theinfluence of changes in gas viscosity and density on thebehaviour of the PNT vary according to precise designand should be both stated by the manufacturer andchecked by the user

Data collection

Calibration of equipment

Equipment calibration has significant influence on thecalculated results and should be performed with the utmostcare and according to the recommendations of the manu-facturer Reliable measurements are unachievable with anunsuitable or defective calibration device It is thereforevital that 1) calibration is performed under identical cir-cumstances to and with the same equipment configurationas during measurements 2) the calibration tools arechecked periodically this requires that any calibratedsyringes or rotameters are returned to the manufacturers ofsuch devices on a regular basis according to the recom-mendations for any specific device (eg 12 monthly forprecision syringes) or whenever any deviation is suspec-ted 3) qualified personnel who understand both the pro-cedure and the signals and parameters displayed performthe calibration4) manual calibration is performed to checkthe automatic calibration procedures and 5) any deviationsin inspired gas viscosity are taken into account in the PNTcalibration

Data acquisition

Data acquisition requirements for infant respiratoryfunction testing are dealt with elsewhere in this series [3]Only those aspects of particular pertinence to tidalbreathing analysis are referred to below As discussedpreviously [3 15] it is crucial that the analogue flowsignal is passed through anti-aliasing filters withappropriate frequency cut-offs prior to sampling in order

1183TIDAL BREATH ANALYSIS IN INFANTS

to satisfy the Shannon sampling theorem and avoidthe potentially insidious problems of aliasing The flowdata are sampled by an analogue-to-digital (AD) con-verter which maps a specified voltage range in to anumber of equally spaced binary numbers It is crucialthat the incoming voltage signal from the flow transduceroccupies as much of the allowable voltage range ofthe AD converter as possible if maximum resolution is tobe attained For example if the flow ranges plusmn30 Lmin-1

and is digitized using a 12-bit AD converter themaximum resolution of the recorded flow signal is 60Lmin-1212 (ie 146 mLmin-1) For this reason togetherwith the need to minimize apparatus dead space andresistance a range of PNTs are probably needed toaccommodate infants of different ages undergoing dif-ferent types of respiratory function test in any one centreThe manufacturer should document both the flow rangeand number of bits of the AD converter

Sampling rate

The necessary sampling rate is determined by Shannonrsquostheorem and the clinical purpose of the tidal breathinganalysis The sampling interval (Dt) between flow datapoints determines the resolution of all identified timepoints such as the beginning and end of inspiration andexpiration Consequently identified time intervals suchas tI and tE have uncertainties of 2Dt For example with anfR of 60 breathsmin-1 and a sampling rate of 100 Hz(Dt=10 ms) the measurement error in tI and tE can be upto 4 A sampling rate of 100 Hz has been shown tobe normally adequate when calculating only VT and fR(see Appendix) whereas greater time resolution may berequired in rapidly breathing infants or for the measure-ment of certain parameters such as tPTEFtE Samplingrates of $200 Hz are therefore recommended foracquisition of tidal breathing data if such analyses are tobe performed particularly in small rapidly breathinginfants [8]

During data collection andor replay time-based dis-plays of flow and volume are required together withsimultaneous displays of flowvolume loops and relevanttrend data [20] These assist in the recognition of air leaksand behavioural state Of particular importance is thedetermination of when the infant has adapted to thepresence of the face mask Exactly how to make thisdetermination remains debatable The decision to com-mence recordings must be based on the operatorrsquos ownexperience plus observation of the displayed signals inorder to ascertain that 1) the breathing pattern is regularstable and representative for that infant 2) there is notrend in instantaneous fR (ie a stable mean fR has beenachieved) and 3) the signals are technically acceptable(eg no leaks artefacts or excessive volume drift)

Once the infant has adapted to the mask and is sleepingquietly and breathing regularly tidal breathing should berecorded in epochs of 30ndash60 s These should be repeatedover the next 5 min and at a later interval if a measure ofreproducibility is desired The number of recorded breat-hing cycles to use for evaluation depends on the variabilityof the signals but should allow the investigator to selectseveral epochs for evaluation It is recommended that eachepoch should contain at least 20 cycles

The essential general information which should berecorded when measuring any lung function parameters ininfants have been described previously [1] For a tidalbreathing study the additional data include time since lastfeeding start and end time of measurement andadaptation time

Signal processing

Numerical integration

Typically the primary measured signal is flow Thismust be integrated with respect to time to produce volumethis is most conveniently performed using a computerThere are a variety of numerical integration methodsavailable all of which connect adjacent data points withsome kind of curve and then sum the calculated areasbeneath each curve segment The more sophisticatedmethods make more accurate interpolations between thedata points than simpler algorithms but at the expense ofgreater complexity One of the simplest numerical inte-gration methods is the so-called trapezoidal rule (fig 2)This assumes that the sampled data points of the flowsignal are connected by straight lines and that the volumeincrement between the ith and (i-1)th data points (DVi) isgiven by

DVi = Dt (Vrsquoi + Vrsquoi 1)=2 (1)

where Vrsquoi and Vrsquoi-1 are the flow at the ith and (i-1)th datapoints The DVi are then summed to yield the total areaunder the curve

V = V0+Xn

i= 1

DVi (2)

where V0 is the volume at which integration of flow begins(which would normally be zero if integration begins at thestart of inspiration)

Flow

Time

Dt

V rsquoi-1V rsquoi

V rsquoi+1 V rsquoi+2

Fig 2 ndash Illustration of the trapezoidal rule The flow signal (mdash) isintegrated by joining its data points using straight lines (- - -) andcalculating the area under each line segment using Equation 1 volume increment between the ith and (i+1)th data points Dt samplinginterval Vrsquoi-1 Vrsquoi Vrsquoi+1 and Vrsquoi+2 flow at (i-1)th ith (i+1)th and (i+2)thdata points

1184 JHT BATES ET AL

Numerical integration is always in error when theoriginal continuous curve is represented by some kindof approximation function between the sampled pointsHowever these errors decrease as the data sampling rateincreases and the sampled points become more closelyspaced For most respiratory applications the integrationerror incurred with the trapezoidal rule is probablynegligible with a data sampling rate of 200 Hz

Volume drift

When flow is integrated to yield volume an upward ordownward drift in the volume baseline is invariably seenSome degree of drift is expected for purely physiologicalreasons For example the respiratory exchange ratio (iecarbon dioxide productionoxygen consumption)is usually~08 ie the volume of O2 absorbed by the lungs is 20greater than the volume of CO2 excreted This is reflectedin a slightly greater volume of gas being inspired thanexpired with each breath Also if the inspired air is notwarmed to body temperature and prehumidified thevolume of gas expired with each breath can be increasedby up to 5 (see discussion of body temperature baro-metric pressure and saturated with water vapour (BTPS)conditions below) relative to that inspired by a gain inwater vapour content These physiological effects con-tribute to a gradually increasing or decreasing volumemeasured at the mouth but not to a real change in baselinelung volume

In addition to the physiological factors discussed abovethe following methodological factors also contribute tovolume drift

Temperature changesbetween inspired and expired gas Ifinspired air is not warmed to body temperature beforepassing through the PNT it has a different viscosity anddensity to expired air which causes the PNT to registerthe transit of an equal number of molecules differentlybetween inspiration and expiration Variations in tem-perature may also affect the physical dimensions of thePNT due to the coefficients of thermal expansion of itscomponents

Changes in gas composition between inspiration andexpiration Inspired and expired gases differ in theirpartial pressures of O2 and CO2 This leads to slightdifferences in the viscosities of the gas mixtures withconcomitant effects on the flows registered duringinspiration and expiration by the PNT

Leaks Any leaks between the airway opening and PNTwhether through the mask seal or around a tracheal tubecause a discrepancy between the volume registered bythe PNT and that entering or leaving the lungs and hence adrift in volume This problem is most likely to occurimmediately after mask displacement if the infant movesor in a pressurized system (eg during artificialventilation)

Zero offset in flow calibration If the true zero flow isregistered as some finite value then integration of thisoffset over time results in a linear drift in volume with aslope equal to the offset Accurate delineation of the zeroflow point is more difficult as the sensitivity of the PNT

decreases which generally occurs as the linear rangeincreases The resolution of the AD converter used tosample the flow also sets a limit on how accurately the zeroflow point can be identified Therefore perfect offsetcompensation is never possible To prevent this volumedrift a dead band around the zero flow in which all valuesare set to zero is used in some devices However a deadband can hamper breath detection especially when flowis very low thus its use and the flow thresholds of thedead band should be described by the manufacturer of theequipment [21]

Imperfections in the pneumotachometer response If thetransducer for measuring flow does not function as aperfect measuring instrument (which is always the case tosome degree and may be significantly so under dynamicconditions) it is unlikely that the inspiratory and expiratoryflows are measured equally This produces asymmetries inthe recorded flow Such asymmetry can often be seen inmeasurements from infants intubated with small endo-tracheal tubes due to the geometric differences on eitherside of the PNT

Correcting volume drift

The analysis of tidal breathing data requires theexamination of data records containing a substantialnumber of breaths (typically $20) obtained during regularbreathing In principle it might be possible to avoid driftin volume in such a data record by preconditioning theinspired gas to BTPS conditions continuously monitoringgas partial pressures in both the alveoli and the pulmonaryarterial and venous blood to correct for respiratory ex-change ratios not equal to unity and eliminating all themethodological factors discussed above However this isextremely difficult if not impossible in practice Conse-quently it is never known how much of the baseline driftin volume is due to drift and how much represents a truechange in absolute lung volume Also because the subjectis assumed to be in the physiological steady state whendata are recorded the assumption is generally made thatfunctional residual capacity (FRC) remains more or lessconstant throughout the study period Such a situation isthus forced on the measured volume signal by some kindof drift correction algorithm which first assesses the driftand then removes it This does not of course mean thatFRC must be identical from one breath to the next butmerely that there is no net upward or downward trend inFRC over a period containing many breaths

Off-line drift correction algorithms commonly definethe drift in volume as the slope of the straight line fittedto the end-expiratory points in an epoch of tidal breaths(fig 3) In order to avoid any outliers skewing the reg-ression it may be useful to exclude those end-expiratorypoints with the greatest deviations and then refit the lineThis line is then subtracted from the volume to removethe drift and the mean level of the new end-expiratorypoints adjusted to zero There are also other ways inwhich volume can be drift-corrected such as subtractionof a curvilinear baseline instead of a straight line orrezeroing of volume at the end of every breath (whichrequires breath detection see below) Different dedrif-ting algorithms usually lead to slight differences in the

1185TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

equipment and performing tests It is anticipated thatacceptance and application of these recommendationswill be of particular value when attempting to comparedata between centres develop or use reference data orparticipate in multicentre trials which use tidal breathingparameters as outcome measures

An infantrsquos breathing pattern measured during tidalbreathing contains significant physiological informationpertaining to a number of processes related to respiratorycontrol and pulmonary mechanical function Such in-formation is encapsulated within a number of conceptually

straightforward parameters The most fundamental para-meters contained in the flow and volume signal tidalvolume (VT) respiratory frequency (fR) and inspiratory (tI)and expiratory time (tE) are shown in figure 1a and b fRwhich is calculated as 60(tI+tE) is also referred to asrespiratory rate (RR) These basic parameters can be usedto calculate quantities pertaining to the more detailedaspects of the pattern and magnitude of tidal breathingsuch as the time to peak tidal expiratory flow (tPTEF) as aproportion of tE (tPTEFtE) minute ventilation (VrsquoE) meaninspiratory flow (VTtI) and the duty cycle (tItotal breathtime (ttot)) In addition tidal breathing flowvolume loopsmay be plotted for visual inspection and evaluation ofparameters such as the volume expired up to the time ofpeak expiratory flow (VPTEF) or defined flow in relationto exhaled volume (fig 1c) It may also be useful toderive certain composite parameters such as VPTEFrelated to VT (VPTEFVT) However the purpose of thisdocument is not to provide a theoretical background tothe factors determining tidal breathing patterns duringearly life nor to comment on the potential clinicalsignificance (or irrelevance in some instances) of theinnumerable parameters that can be calculated fromrecordings of tidal breathing For this the reader isdirected to the existing literature much of which hasbeen summarized recently [7] The intent is rather tostress the various factors during data acquisition analysisand reporting of tidal breathing parameters that can leadto systematic errors between different recording systems

Obtaining tidal breathing parameters requires nothingmore than the measurement of flow or volume at the mouthand nose during a period of regular breathing Howeverthere are a number of considerations pertaining to theanalysis of tidal breathing data that make it somewhat lessthan trivial The most problematic issue from a practicalpoint of view is the precise detection of the onset ofinspiration and expiration The problem of automaticallysegmenting breaths into their inspiratory and expiratoryphases is thus considered in the present document Otherpractical issues such as the collection of airflow data andthe derivation of a drift-free volume signal from such dataare also considered Some of these practical issues areinfluenced by the task in hand For example investigationsinto the control of breathing require the determination ofnot only the magnitude of but also the variability in VT tIand tE On the one hand this requires measurement of alarge number of breathing cycles which cannot be per-formed with the conventional combination of face maskand pneumotachometer (PNT) due to their relatively highapparatus dead space thus alternative approaches areusually required [8] On the other hand a relatively lowsample rate (10ndash50 Hz) will often suffice for such studiesBy contrast when attempting to analyse the mechanicalcomponents of the tidal breathing signal a relatively lownumber of consecutive reproducible and artefact-freebreaths need to be recorded but at a higher sampling rate

Because of the complexities of using noninvasive bodysurface measurements to obtain quantitatively accurateassessments of tidal breathing parameters and the lack ofany standardized approach to such measurements thecurrent document is limited to direct measurements ofairflow and volume at the airway opening Neverthelesssome of the recommendations within may be pertinent ifalternative approaches are used

Volu

me

a)

tEtI

Flow

b)

tPTIF

PTIF

PTEF

tPTEF

Time

Insp Exp

ttot

VT

Insp Exp

c)

Flow

Volume

075 05 025 01VT

VPTEF

PTEF TEF50

TIF50PTIF

Insp

Exp

Fig 1 ndash Tidal breathing parameters of a) the flow and b) volumesignals and c) the flowvolume loop The tidal expiratory (TEF50) andinspiratory flow when 50 of tidal volume (VT) remains in the lung(TIF50) are shown other values of TEFx and TIFx correspond to thedifferent VT on the scale bar Insp inspiration exp expiration tIinspiratory time tE expiratory time ttot total breath time PTIF peaktidal inspiratory flow PTEF peak tidal expiratory flow tPTEF time topeak expiratory flow tPTIF time to peak inspiratory flow VPTEFvolume expired before PTEF attained

1181TIDAL BREATH ANALYSIS IN INFANTS

General procedures

In the present paper tidal breathing is taken to be thenatural physiological state of undisturbed regular breath-ing Measurements are commonly performed during quietsleep which in older infants is usually associated withactive inspiration and passive expiration In order to recordperiods of undisturbed breathing there should be neg-ligible influence of the measuring equipment [9] thustidal breathing should be assessed prior to more complexinvestigations For example although it is certainly pos-sible to measure tidal breathing parameters immediatelyafter insertion of an oesophageal catheter the valuesobtained will be significantly different from those mea-sured prior to catheter insertion Meaningful comparisonof tidal breathing parameters within and between infantsor interpretation with respect to reference values areonly possible if the measurement conditionsare the sameCareful documentation of measurement conditions andequipment characteristics during tidal breathing measure-ments is therefore essential

Contrary to common belief collection of undisturbedquiet tidal breathing data is not a trivial undertaking aseven the application of a face mask can stimulate breath-ing [10 11] Consequently measurements should onlybe made after a sufficient adaptation period has beenallowed to elapse after attachment of the measurementequipment This is not a problem if there is negligibleadditional dead space involved such as with the flow-through technique [8] However when using a conven-tional mask and PNT the apparatus dead space limits theduration of measurements [12] Furthermore the smallerthe infant the greater the impact of the apparatus deadspace Careful preparation is therefore necessary tominimize the adaptation period especially when studyingnewborns

Careful use of equipment in order to ensure patientsafety remains the responsibility of the operator [1]Routine safety measures in the pulmonary functiontesting laboratory include 1) availability of full resuscita-tion equipment including suction at the site of infantlung function testing plus a suitable alarm system 2)the presence of two individuals (other than parents)during testing one of whom has prime responsibility forthe infantrsquos wellbeing the infant must never be leftunattended 3) continuous physiological monitoring (andideally recording) including at least pulse oximetry 4)use of transparent face masks and 5) adherence to thehospital-specific protocol for sedation or anaesthesia

Further details regarding measurement conditionswhich may influence infant safety or the accuracy andreproducibility of results have been published previously[5 13] The measurement procedure can be summarizedas follows 1) The infant should be lying supine with theneck andor shoulders supported in the midline in slightextension This position should be stabilized by use of aneck roll or head ring If an alternative posture is usedthis should be stated clearly in any publication 2) Theinfant should be fed dry clean and comfortably clothed3) The face mask should cover mouth and nose and beplaced with minimal pressure while also being airtightTo prevent any air leaks a thin sealing ring of siliconputty is often helpful

Equipment

Airflow and volume determined by numerical integra-tion of airflow are the basic signals of tidal breathinganalysis (fig 1) Airflow is usually measured at the airwayopening using a PNT connected to a face mask Althoughboth respiratory inductance and face-out body plethys-mography have been used to record tidal breathingsignals the former only provides quantitatively accuratemeasurements if there has been previous calibration usinga PNT and the measurement conditions remain verystable [14] whereas the latter is too expensive andcumbersome for routine bedside application A detaileddescription of devices available for the measurement offlow and hence volume has been published recently[15] The reader is also referred to the document in thepresent series on specifications for equipment and soft-ware used for infant pulmonary function testing forfurther details and justification of the recommendationspresented below [1 3]

Face mask

Different types of mask are used based largely on localpreference and availability The dead space of the facemask and flowmeter may have a marked influence onthe pattern of breathing with respect to both magnitudeand timing thereby altering the very parameters that theinvestigator is trying to assess In preterm and sickneonates even when a small face mask and a very-low-dead space flowmeter are used the total apparatus deadspace may exceed the infantrsquos own dead space and pre-sent a considerable load thereby precluding anything butbrief intermittent recordings In such infants the use of thedead-space-free-flow-through technique is recommended[8]

In order to aid international standardization thefollowing recommendations are made 1) report brandand size of mask and whether the mask has an air-filledcuff since the mask dead space depends significantly onthe cuff pressure 2) document other aids to improvingairtight placement (eg silicon putty or Vaseline) 3) mini-mize dead space and report whether measures to reducemask volume are used (eg shortening of the mask port orputting silicon putty in the mask) and 4) measure the deadspace of the mask by water displacement and subtract 50of this volume to estimate the effective dead space [1 16]

Flowmeter

The conventional transducer employed for measurementof respiratory flow is the combination of PNT and diffe-rential pressure transducer Attention is restricted to thisdevice in the present document although several newdevices such as ultrasonic flowmeters and hot wireanemometers [17] are currently being developed andvalidated for use in infant respiratory function measure-ments A number of general considerations concerninginstrumentation and measurement technique pertain tothe measurement of flow and its recording using acomputer These are as follows 1) The flowmeter shouldbe a low-resistance low-dead-space device Flowmeters

1182 JHT BATES ET AL

for measurements in preterm infants should have a deadspace of lt15 mL Unfortunately the apparatus deadspace of most modern devices arises largely due to thenecessary connecting ports and so possibilities for deadspace reduction are limited In any case the connectionbetween PNT and mask must be minimized withoutdisturbing the linearity of the flowmeter [15] 2) Mini-mizing the resistance of the infant lung function equip-ment is important since the overall resistance of theequipment may not only dramatically change the respi-ratory pattern in spontaneously breathing babies butalso interfere with triggering devices in those who areventilated Any significant increase in resistance in-creases the expiratory time constant and potentiallyinfluences the end-expiratory level This in turn affectsany measurements that are volume-dependent includingvarious tidal breathing parameters The need to designfuture apparatus with as low a resistance as possiblewithin the constraints of simultaneously attaining a lowdead space and high resolution cannot be overempha-sized 3) The combined resistance of the apparatus (in-cluding any valves capnographs etc) should be lt20of the infantrsquos intrinsic resistance at the mean flowslikely to be encountered [1] Thus as a rough guide thecombined apparatus resistance should not exceed 12kPaL-1s at 50 mLs-1 in spontaneouslybreathing preterminfants 07 kPaL-1s at 100 mLs-1 in term neonates and05 kPaL-1s at 300 mLs-1 in infants and young children4) The response of the flowmeter should be linear overthe range of flows encountered The extent to which aPNT remains linear over an extended range is criticallydependent on design features such as whether it is acapillary- screen- or variable orifice-type device and thegeometry of any integral connections It is thereforeessential that the manufacturer provides accurate detailsand that the user checks the range of flows over whichthe flowmeter provides accurate recordings The approx-imate linearity ranges required for various sizes of infantare 0ndash100 mLs-1 in preterm infants and neonates of 2ndash4kg 0ndash300 mLs-1 in infants of 4ndash10 kg and 0ndash500 mLs-1

in preschool children of 11ndash15 kg In practice flowmeterswith a linear range of 0ndash10 Lmin-1 are commonly used inpreterm infants and neonates whereas those with a rangeof 0ndash35 Lmin-1 are used for obtaining tidal breathingmeasurements in older infants and young children 5) If aPNT has a nonlinear response over the desired flowrange it may be possible to effectively linearize it bycharacterizing the response and inverting it digitally Thiscan reduce dead space by allowing the use of a smaller-calibre PNT However it must be ensured that theresponse characteristics of the device remains constantover a prolonged period after repeated disinfection andon exposure to different respired gases 6) The flowmetershould have a flat frequency response over a frequencyrange sufficient to encompass the majority of the power inthe measured signals [1] For tidal breathing signals it isprobably sufficient to have a flat frequency response upto 10 Hz If the transducer itself does not have a flatfrequency response over this range it may be possibleto render it flat by digital compensation of the sampleddata [18] however this is only possible if the responseof the device is linear 7) If a PNT with metal screensor capillaries is used it should be heated to body tem-perature to avoid condensation on the resistive element

Major changes in screen resistance and hence measuredflows can occur within lt1 min of placing an unheatedPNT into a ventilator circuit thus this practice is stronglydiscouraged 8) Despite such heating PNTs with screensor capillaries in the ventilator circuit are highly suscep-tible to obstruction by secretions leading to falsely highmeasured flows and possible danger to the patientTherefore these PNTs should only be used by qualifiedpersonnel while the patient is under direct observation[19] 9) The geometry of the connectors on either side ofthe PNT screen affects the overall pressureflow charac-teristics of the device It is therefore important that theconnectors be as symmetrical as possible on either side ofthe PNT and that the PNT is calibrated in situ in exactlythe same configuration as that to be used with the subject[15] 10) If the inspired gas differs significantly fromroom air (eg by increased inspiratory oxygen fraction(FIO2)) it may be of different viscosity to room air andtherefore have different PNT calibration factors In such acase either the PNT should be calibrated with the inspiredgas or the room air calibration factors should be scaled bythe relative differences in gas viscosity [15] For measure-ments during artificial ventilation continuous FIO2correction at the bedside is advantageous [19] Theinfluence of changes in gas viscosity and density on thebehaviour of the PNT vary according to precise designand should be both stated by the manufacturer andchecked by the user

Data collection

Calibration of equipment

Equipment calibration has significant influence on thecalculated results and should be performed with the utmostcare and according to the recommendations of the manu-facturer Reliable measurements are unachievable with anunsuitable or defective calibration device It is thereforevital that 1) calibration is performed under identical cir-cumstances to and with the same equipment configurationas during measurements 2) the calibration tools arechecked periodically this requires that any calibratedsyringes or rotameters are returned to the manufacturers ofsuch devices on a regular basis according to the recom-mendations for any specific device (eg 12 monthly forprecision syringes) or whenever any deviation is suspec-ted 3) qualified personnel who understand both the pro-cedure and the signals and parameters displayed performthe calibration4) manual calibration is performed to checkthe automatic calibration procedures and 5) any deviationsin inspired gas viscosity are taken into account in the PNTcalibration

Data acquisition

Data acquisition requirements for infant respiratoryfunction testing are dealt with elsewhere in this series [3]Only those aspects of particular pertinence to tidalbreathing analysis are referred to below As discussedpreviously [3 15] it is crucial that the analogue flowsignal is passed through anti-aliasing filters withappropriate frequency cut-offs prior to sampling in order

1183TIDAL BREATH ANALYSIS IN INFANTS

to satisfy the Shannon sampling theorem and avoidthe potentially insidious problems of aliasing The flowdata are sampled by an analogue-to-digital (AD) con-verter which maps a specified voltage range in to anumber of equally spaced binary numbers It is crucialthat the incoming voltage signal from the flow transduceroccupies as much of the allowable voltage range ofthe AD converter as possible if maximum resolution is tobe attained For example if the flow ranges plusmn30 Lmin-1

and is digitized using a 12-bit AD converter themaximum resolution of the recorded flow signal is 60Lmin-1212 (ie 146 mLmin-1) For this reason togetherwith the need to minimize apparatus dead space andresistance a range of PNTs are probably needed toaccommodate infants of different ages undergoing dif-ferent types of respiratory function test in any one centreThe manufacturer should document both the flow rangeand number of bits of the AD converter

Sampling rate

The necessary sampling rate is determined by Shannonrsquostheorem and the clinical purpose of the tidal breathinganalysis The sampling interval (Dt) between flow datapoints determines the resolution of all identified timepoints such as the beginning and end of inspiration andexpiration Consequently identified time intervals suchas tI and tE have uncertainties of 2Dt For example with anfR of 60 breathsmin-1 and a sampling rate of 100 Hz(Dt=10 ms) the measurement error in tI and tE can be upto 4 A sampling rate of 100 Hz has been shown tobe normally adequate when calculating only VT and fR(see Appendix) whereas greater time resolution may berequired in rapidly breathing infants or for the measure-ment of certain parameters such as tPTEFtE Samplingrates of $200 Hz are therefore recommended foracquisition of tidal breathing data if such analyses are tobe performed particularly in small rapidly breathinginfants [8]

During data collection andor replay time-based dis-plays of flow and volume are required together withsimultaneous displays of flowvolume loops and relevanttrend data [20] These assist in the recognition of air leaksand behavioural state Of particular importance is thedetermination of when the infant has adapted to thepresence of the face mask Exactly how to make thisdetermination remains debatable The decision to com-mence recordings must be based on the operatorrsquos ownexperience plus observation of the displayed signals inorder to ascertain that 1) the breathing pattern is regularstable and representative for that infant 2) there is notrend in instantaneous fR (ie a stable mean fR has beenachieved) and 3) the signals are technically acceptable(eg no leaks artefacts or excessive volume drift)

Once the infant has adapted to the mask and is sleepingquietly and breathing regularly tidal breathing should berecorded in epochs of 30ndash60 s These should be repeatedover the next 5 min and at a later interval if a measure ofreproducibility is desired The number of recorded breat-hing cycles to use for evaluation depends on the variabilityof the signals but should allow the investigator to selectseveral epochs for evaluation It is recommended that eachepoch should contain at least 20 cycles

The essential general information which should berecorded when measuring any lung function parameters ininfants have been described previously [1] For a tidalbreathing study the additional data include time since lastfeeding start and end time of measurement andadaptation time

Signal processing

Numerical integration

Typically the primary measured signal is flow Thismust be integrated with respect to time to produce volumethis is most conveniently performed using a computerThere are a variety of numerical integration methodsavailable all of which connect adjacent data points withsome kind of curve and then sum the calculated areasbeneath each curve segment The more sophisticatedmethods make more accurate interpolations between thedata points than simpler algorithms but at the expense ofgreater complexity One of the simplest numerical inte-gration methods is the so-called trapezoidal rule (fig 2)This assumes that the sampled data points of the flowsignal are connected by straight lines and that the volumeincrement between the ith and (i-1)th data points (DVi) isgiven by

DVi = Dt (Vrsquoi + Vrsquoi 1)=2 (1)

where Vrsquoi and Vrsquoi-1 are the flow at the ith and (i-1)th datapoints The DVi are then summed to yield the total areaunder the curve

V = V0+Xn

i= 1

DVi (2)

where V0 is the volume at which integration of flow begins(which would normally be zero if integration begins at thestart of inspiration)

Flow

Time

Dt

V rsquoi-1V rsquoi

V rsquoi+1 V rsquoi+2

Fig 2 ndash Illustration of the trapezoidal rule The flow signal (mdash) isintegrated by joining its data points using straight lines (- - -) andcalculating the area under each line segment using Equation 1 volume increment between the ith and (i+1)th data points Dt samplinginterval Vrsquoi-1 Vrsquoi Vrsquoi+1 and Vrsquoi+2 flow at (i-1)th ith (i+1)th and (i+2)thdata points

1184 JHT BATES ET AL

Numerical integration is always in error when theoriginal continuous curve is represented by some kindof approximation function between the sampled pointsHowever these errors decrease as the data sampling rateincreases and the sampled points become more closelyspaced For most respiratory applications the integrationerror incurred with the trapezoidal rule is probablynegligible with a data sampling rate of 200 Hz

Volume drift

When flow is integrated to yield volume an upward ordownward drift in the volume baseline is invariably seenSome degree of drift is expected for purely physiologicalreasons For example the respiratory exchange ratio (iecarbon dioxide productionoxygen consumption)is usually~08 ie the volume of O2 absorbed by the lungs is 20greater than the volume of CO2 excreted This is reflectedin a slightly greater volume of gas being inspired thanexpired with each breath Also if the inspired air is notwarmed to body temperature and prehumidified thevolume of gas expired with each breath can be increasedby up to 5 (see discussion of body temperature baro-metric pressure and saturated with water vapour (BTPS)conditions below) relative to that inspired by a gain inwater vapour content These physiological effects con-tribute to a gradually increasing or decreasing volumemeasured at the mouth but not to a real change in baselinelung volume

In addition to the physiological factors discussed abovethe following methodological factors also contribute tovolume drift

Temperature changesbetween inspired and expired gas Ifinspired air is not warmed to body temperature beforepassing through the PNT it has a different viscosity anddensity to expired air which causes the PNT to registerthe transit of an equal number of molecules differentlybetween inspiration and expiration Variations in tem-perature may also affect the physical dimensions of thePNT due to the coefficients of thermal expansion of itscomponents

Changes in gas composition between inspiration andexpiration Inspired and expired gases differ in theirpartial pressures of O2 and CO2 This leads to slightdifferences in the viscosities of the gas mixtures withconcomitant effects on the flows registered duringinspiration and expiration by the PNT

Leaks Any leaks between the airway opening and PNTwhether through the mask seal or around a tracheal tubecause a discrepancy between the volume registered bythe PNT and that entering or leaving the lungs and hence adrift in volume This problem is most likely to occurimmediately after mask displacement if the infant movesor in a pressurized system (eg during artificialventilation)

Zero offset in flow calibration If the true zero flow isregistered as some finite value then integration of thisoffset over time results in a linear drift in volume with aslope equal to the offset Accurate delineation of the zeroflow point is more difficult as the sensitivity of the PNT

decreases which generally occurs as the linear rangeincreases The resolution of the AD converter used tosample the flow also sets a limit on how accurately the zeroflow point can be identified Therefore perfect offsetcompensation is never possible To prevent this volumedrift a dead band around the zero flow in which all valuesare set to zero is used in some devices However a deadband can hamper breath detection especially when flowis very low thus its use and the flow thresholds of thedead band should be described by the manufacturer of theequipment [21]

Imperfections in the pneumotachometer response If thetransducer for measuring flow does not function as aperfect measuring instrument (which is always the case tosome degree and may be significantly so under dynamicconditions) it is unlikely that the inspiratory and expiratoryflows are measured equally This produces asymmetries inthe recorded flow Such asymmetry can often be seen inmeasurements from infants intubated with small endo-tracheal tubes due to the geometric differences on eitherside of the PNT

Correcting volume drift

The analysis of tidal breathing data requires theexamination of data records containing a substantialnumber of breaths (typically $20) obtained during regularbreathing In principle it might be possible to avoid driftin volume in such a data record by preconditioning theinspired gas to BTPS conditions continuously monitoringgas partial pressures in both the alveoli and the pulmonaryarterial and venous blood to correct for respiratory ex-change ratios not equal to unity and eliminating all themethodological factors discussed above However this isextremely difficult if not impossible in practice Conse-quently it is never known how much of the baseline driftin volume is due to drift and how much represents a truechange in absolute lung volume Also because the subjectis assumed to be in the physiological steady state whendata are recorded the assumption is generally made thatfunctional residual capacity (FRC) remains more or lessconstant throughout the study period Such a situation isthus forced on the measured volume signal by some kindof drift correction algorithm which first assesses the driftand then removes it This does not of course mean thatFRC must be identical from one breath to the next butmerely that there is no net upward or downward trend inFRC over a period containing many breaths

Off-line drift correction algorithms commonly definethe drift in volume as the slope of the straight line fittedto the end-expiratory points in an epoch of tidal breaths(fig 3) In order to avoid any outliers skewing the reg-ression it may be useful to exclude those end-expiratorypoints with the greatest deviations and then refit the lineThis line is then subtracted from the volume to removethe drift and the mean level of the new end-expiratorypoints adjusted to zero There are also other ways inwhich volume can be drift-corrected such as subtractionof a curvilinear baseline instead of a straight line orrezeroing of volume at the end of every breath (whichrequires breath detection see below) Different dedrif-ting algorithms usually lead to slight differences in the

1185TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

General procedures

In the present paper tidal breathing is taken to be thenatural physiological state of undisturbed regular breath-ing Measurements are commonly performed during quietsleep which in older infants is usually associated withactive inspiration and passive expiration In order to recordperiods of undisturbed breathing there should be neg-ligible influence of the measuring equipment [9] thustidal breathing should be assessed prior to more complexinvestigations For example although it is certainly pos-sible to measure tidal breathing parameters immediatelyafter insertion of an oesophageal catheter the valuesobtained will be significantly different from those mea-sured prior to catheter insertion Meaningful comparisonof tidal breathing parameters within and between infantsor interpretation with respect to reference values areonly possible if the measurement conditionsare the sameCareful documentation of measurement conditions andequipment characteristics during tidal breathing measure-ments is therefore essential

Contrary to common belief collection of undisturbedquiet tidal breathing data is not a trivial undertaking aseven the application of a face mask can stimulate breath-ing [10 11] Consequently measurements should onlybe made after a sufficient adaptation period has beenallowed to elapse after attachment of the measurementequipment This is not a problem if there is negligibleadditional dead space involved such as with the flow-through technique [8] However when using a conven-tional mask and PNT the apparatus dead space limits theduration of measurements [12] Furthermore the smallerthe infant the greater the impact of the apparatus deadspace Careful preparation is therefore necessary tominimize the adaptation period especially when studyingnewborns

Careful use of equipment in order to ensure patientsafety remains the responsibility of the operator [1]Routine safety measures in the pulmonary functiontesting laboratory include 1) availability of full resuscita-tion equipment including suction at the site of infantlung function testing plus a suitable alarm system 2)the presence of two individuals (other than parents)during testing one of whom has prime responsibility forthe infantrsquos wellbeing the infant must never be leftunattended 3) continuous physiological monitoring (andideally recording) including at least pulse oximetry 4)use of transparent face masks and 5) adherence to thehospital-specific protocol for sedation or anaesthesia

Further details regarding measurement conditionswhich may influence infant safety or the accuracy andreproducibility of results have been published previously[5 13] The measurement procedure can be summarizedas follows 1) The infant should be lying supine with theneck andor shoulders supported in the midline in slightextension This position should be stabilized by use of aneck roll or head ring If an alternative posture is usedthis should be stated clearly in any publication 2) Theinfant should be fed dry clean and comfortably clothed3) The face mask should cover mouth and nose and beplaced with minimal pressure while also being airtightTo prevent any air leaks a thin sealing ring of siliconputty is often helpful

Equipment

Airflow and volume determined by numerical integra-tion of airflow are the basic signals of tidal breathinganalysis (fig 1) Airflow is usually measured at the airwayopening using a PNT connected to a face mask Althoughboth respiratory inductance and face-out body plethys-mography have been used to record tidal breathingsignals the former only provides quantitatively accuratemeasurements if there has been previous calibration usinga PNT and the measurement conditions remain verystable [14] whereas the latter is too expensive andcumbersome for routine bedside application A detaileddescription of devices available for the measurement offlow and hence volume has been published recently[15] The reader is also referred to the document in thepresent series on specifications for equipment and soft-ware used for infant pulmonary function testing forfurther details and justification of the recommendationspresented below [1 3]

Face mask

Different types of mask are used based largely on localpreference and availability The dead space of the facemask and flowmeter may have a marked influence onthe pattern of breathing with respect to both magnitudeand timing thereby altering the very parameters that theinvestigator is trying to assess In preterm and sickneonates even when a small face mask and a very-low-dead space flowmeter are used the total apparatus deadspace may exceed the infantrsquos own dead space and pre-sent a considerable load thereby precluding anything butbrief intermittent recordings In such infants the use of thedead-space-free-flow-through technique is recommended[8]

In order to aid international standardization thefollowing recommendations are made 1) report brandand size of mask and whether the mask has an air-filledcuff since the mask dead space depends significantly onthe cuff pressure 2) document other aids to improvingairtight placement (eg silicon putty or Vaseline) 3) mini-mize dead space and report whether measures to reducemask volume are used (eg shortening of the mask port orputting silicon putty in the mask) and 4) measure the deadspace of the mask by water displacement and subtract 50of this volume to estimate the effective dead space [1 16]

Flowmeter

The conventional transducer employed for measurementof respiratory flow is the combination of PNT and diffe-rential pressure transducer Attention is restricted to thisdevice in the present document although several newdevices such as ultrasonic flowmeters and hot wireanemometers [17] are currently being developed andvalidated for use in infant respiratory function measure-ments A number of general considerations concerninginstrumentation and measurement technique pertain tothe measurement of flow and its recording using acomputer These are as follows 1) The flowmeter shouldbe a low-resistance low-dead-space device Flowmeters

1182 JHT BATES ET AL

for measurements in preterm infants should have a deadspace of lt15 mL Unfortunately the apparatus deadspace of most modern devices arises largely due to thenecessary connecting ports and so possibilities for deadspace reduction are limited In any case the connectionbetween PNT and mask must be minimized withoutdisturbing the linearity of the flowmeter [15] 2) Mini-mizing the resistance of the infant lung function equip-ment is important since the overall resistance of theequipment may not only dramatically change the respi-ratory pattern in spontaneously breathing babies butalso interfere with triggering devices in those who areventilated Any significant increase in resistance in-creases the expiratory time constant and potentiallyinfluences the end-expiratory level This in turn affectsany measurements that are volume-dependent includingvarious tidal breathing parameters The need to designfuture apparatus with as low a resistance as possiblewithin the constraints of simultaneously attaining a lowdead space and high resolution cannot be overempha-sized 3) The combined resistance of the apparatus (in-cluding any valves capnographs etc) should be lt20of the infantrsquos intrinsic resistance at the mean flowslikely to be encountered [1] Thus as a rough guide thecombined apparatus resistance should not exceed 12kPaL-1s at 50 mLs-1 in spontaneouslybreathing preterminfants 07 kPaL-1s at 100 mLs-1 in term neonates and05 kPaL-1s at 300 mLs-1 in infants and young children4) The response of the flowmeter should be linear overthe range of flows encountered The extent to which aPNT remains linear over an extended range is criticallydependent on design features such as whether it is acapillary- screen- or variable orifice-type device and thegeometry of any integral connections It is thereforeessential that the manufacturer provides accurate detailsand that the user checks the range of flows over whichthe flowmeter provides accurate recordings The approx-imate linearity ranges required for various sizes of infantare 0ndash100 mLs-1 in preterm infants and neonates of 2ndash4kg 0ndash300 mLs-1 in infants of 4ndash10 kg and 0ndash500 mLs-1

in preschool children of 11ndash15 kg In practice flowmeterswith a linear range of 0ndash10 Lmin-1 are commonly used inpreterm infants and neonates whereas those with a rangeof 0ndash35 Lmin-1 are used for obtaining tidal breathingmeasurements in older infants and young children 5) If aPNT has a nonlinear response over the desired flowrange it may be possible to effectively linearize it bycharacterizing the response and inverting it digitally Thiscan reduce dead space by allowing the use of a smaller-calibre PNT However it must be ensured that theresponse characteristics of the device remains constantover a prolonged period after repeated disinfection andon exposure to different respired gases 6) The flowmetershould have a flat frequency response over a frequencyrange sufficient to encompass the majority of the power inthe measured signals [1] For tidal breathing signals it isprobably sufficient to have a flat frequency response upto 10 Hz If the transducer itself does not have a flatfrequency response over this range it may be possibleto render it flat by digital compensation of the sampleddata [18] however this is only possible if the responseof the device is linear 7) If a PNT with metal screensor capillaries is used it should be heated to body tem-perature to avoid condensation on the resistive element

Major changes in screen resistance and hence measuredflows can occur within lt1 min of placing an unheatedPNT into a ventilator circuit thus this practice is stronglydiscouraged 8) Despite such heating PNTs with screensor capillaries in the ventilator circuit are highly suscep-tible to obstruction by secretions leading to falsely highmeasured flows and possible danger to the patientTherefore these PNTs should only be used by qualifiedpersonnel while the patient is under direct observation[19] 9) The geometry of the connectors on either side ofthe PNT screen affects the overall pressureflow charac-teristics of the device It is therefore important that theconnectors be as symmetrical as possible on either side ofthe PNT and that the PNT is calibrated in situ in exactlythe same configuration as that to be used with the subject[15] 10) If the inspired gas differs significantly fromroom air (eg by increased inspiratory oxygen fraction(FIO2)) it may be of different viscosity to room air andtherefore have different PNT calibration factors In such acase either the PNT should be calibrated with the inspiredgas or the room air calibration factors should be scaled bythe relative differences in gas viscosity [15] For measure-ments during artificial ventilation continuous FIO2correction at the bedside is advantageous [19] Theinfluence of changes in gas viscosity and density on thebehaviour of the PNT vary according to precise designand should be both stated by the manufacturer andchecked by the user

Data collection

Calibration of equipment

Equipment calibration has significant influence on thecalculated results and should be performed with the utmostcare and according to the recommendations of the manu-facturer Reliable measurements are unachievable with anunsuitable or defective calibration device It is thereforevital that 1) calibration is performed under identical cir-cumstances to and with the same equipment configurationas during measurements 2) the calibration tools arechecked periodically this requires that any calibratedsyringes or rotameters are returned to the manufacturers ofsuch devices on a regular basis according to the recom-mendations for any specific device (eg 12 monthly forprecision syringes) or whenever any deviation is suspec-ted 3) qualified personnel who understand both the pro-cedure and the signals and parameters displayed performthe calibration4) manual calibration is performed to checkthe automatic calibration procedures and 5) any deviationsin inspired gas viscosity are taken into account in the PNTcalibration

Data acquisition

Data acquisition requirements for infant respiratoryfunction testing are dealt with elsewhere in this series [3]Only those aspects of particular pertinence to tidalbreathing analysis are referred to below As discussedpreviously [3 15] it is crucial that the analogue flowsignal is passed through anti-aliasing filters withappropriate frequency cut-offs prior to sampling in order

1183TIDAL BREATH ANALYSIS IN INFANTS

to satisfy the Shannon sampling theorem and avoidthe potentially insidious problems of aliasing The flowdata are sampled by an analogue-to-digital (AD) con-verter which maps a specified voltage range in to anumber of equally spaced binary numbers It is crucialthat the incoming voltage signal from the flow transduceroccupies as much of the allowable voltage range ofthe AD converter as possible if maximum resolution is tobe attained For example if the flow ranges plusmn30 Lmin-1

and is digitized using a 12-bit AD converter themaximum resolution of the recorded flow signal is 60Lmin-1212 (ie 146 mLmin-1) For this reason togetherwith the need to minimize apparatus dead space andresistance a range of PNTs are probably needed toaccommodate infants of different ages undergoing dif-ferent types of respiratory function test in any one centreThe manufacturer should document both the flow rangeand number of bits of the AD converter

Sampling rate

The necessary sampling rate is determined by Shannonrsquostheorem and the clinical purpose of the tidal breathinganalysis The sampling interval (Dt) between flow datapoints determines the resolution of all identified timepoints such as the beginning and end of inspiration andexpiration Consequently identified time intervals suchas tI and tE have uncertainties of 2Dt For example with anfR of 60 breathsmin-1 and a sampling rate of 100 Hz(Dt=10 ms) the measurement error in tI and tE can be upto 4 A sampling rate of 100 Hz has been shown tobe normally adequate when calculating only VT and fR(see Appendix) whereas greater time resolution may berequired in rapidly breathing infants or for the measure-ment of certain parameters such as tPTEFtE Samplingrates of $200 Hz are therefore recommended foracquisition of tidal breathing data if such analyses are tobe performed particularly in small rapidly breathinginfants [8]

During data collection andor replay time-based dis-plays of flow and volume are required together withsimultaneous displays of flowvolume loops and relevanttrend data [20] These assist in the recognition of air leaksand behavioural state Of particular importance is thedetermination of when the infant has adapted to thepresence of the face mask Exactly how to make thisdetermination remains debatable The decision to com-mence recordings must be based on the operatorrsquos ownexperience plus observation of the displayed signals inorder to ascertain that 1) the breathing pattern is regularstable and representative for that infant 2) there is notrend in instantaneous fR (ie a stable mean fR has beenachieved) and 3) the signals are technically acceptable(eg no leaks artefacts or excessive volume drift)

Once the infant has adapted to the mask and is sleepingquietly and breathing regularly tidal breathing should berecorded in epochs of 30ndash60 s These should be repeatedover the next 5 min and at a later interval if a measure ofreproducibility is desired The number of recorded breat-hing cycles to use for evaluation depends on the variabilityof the signals but should allow the investigator to selectseveral epochs for evaluation It is recommended that eachepoch should contain at least 20 cycles

The essential general information which should berecorded when measuring any lung function parameters ininfants have been described previously [1] For a tidalbreathing study the additional data include time since lastfeeding start and end time of measurement andadaptation time

Signal processing

Numerical integration

Typically the primary measured signal is flow Thismust be integrated with respect to time to produce volumethis is most conveniently performed using a computerThere are a variety of numerical integration methodsavailable all of which connect adjacent data points withsome kind of curve and then sum the calculated areasbeneath each curve segment The more sophisticatedmethods make more accurate interpolations between thedata points than simpler algorithms but at the expense ofgreater complexity One of the simplest numerical inte-gration methods is the so-called trapezoidal rule (fig 2)This assumes that the sampled data points of the flowsignal are connected by straight lines and that the volumeincrement between the ith and (i-1)th data points (DVi) isgiven by

DVi = Dt (Vrsquoi + Vrsquoi 1)=2 (1)

where Vrsquoi and Vrsquoi-1 are the flow at the ith and (i-1)th datapoints The DVi are then summed to yield the total areaunder the curve

V = V0+Xn

i= 1

DVi (2)

where V0 is the volume at which integration of flow begins(which would normally be zero if integration begins at thestart of inspiration)

Flow

Time

Dt

V rsquoi-1V rsquoi

V rsquoi+1 V rsquoi+2

Fig 2 ndash Illustration of the trapezoidal rule The flow signal (mdash) isintegrated by joining its data points using straight lines (- - -) andcalculating the area under each line segment using Equation 1 volume increment between the ith and (i+1)th data points Dt samplinginterval Vrsquoi-1 Vrsquoi Vrsquoi+1 and Vrsquoi+2 flow at (i-1)th ith (i+1)th and (i+2)thdata points

1184 JHT BATES ET AL

Numerical integration is always in error when theoriginal continuous curve is represented by some kindof approximation function between the sampled pointsHowever these errors decrease as the data sampling rateincreases and the sampled points become more closelyspaced For most respiratory applications the integrationerror incurred with the trapezoidal rule is probablynegligible with a data sampling rate of 200 Hz

Volume drift

When flow is integrated to yield volume an upward ordownward drift in the volume baseline is invariably seenSome degree of drift is expected for purely physiologicalreasons For example the respiratory exchange ratio (iecarbon dioxide productionoxygen consumption)is usually~08 ie the volume of O2 absorbed by the lungs is 20greater than the volume of CO2 excreted This is reflectedin a slightly greater volume of gas being inspired thanexpired with each breath Also if the inspired air is notwarmed to body temperature and prehumidified thevolume of gas expired with each breath can be increasedby up to 5 (see discussion of body temperature baro-metric pressure and saturated with water vapour (BTPS)conditions below) relative to that inspired by a gain inwater vapour content These physiological effects con-tribute to a gradually increasing or decreasing volumemeasured at the mouth but not to a real change in baselinelung volume

In addition to the physiological factors discussed abovethe following methodological factors also contribute tovolume drift

Temperature changesbetween inspired and expired gas Ifinspired air is not warmed to body temperature beforepassing through the PNT it has a different viscosity anddensity to expired air which causes the PNT to registerthe transit of an equal number of molecules differentlybetween inspiration and expiration Variations in tem-perature may also affect the physical dimensions of thePNT due to the coefficients of thermal expansion of itscomponents

Changes in gas composition between inspiration andexpiration Inspired and expired gases differ in theirpartial pressures of O2 and CO2 This leads to slightdifferences in the viscosities of the gas mixtures withconcomitant effects on the flows registered duringinspiration and expiration by the PNT

Leaks Any leaks between the airway opening and PNTwhether through the mask seal or around a tracheal tubecause a discrepancy between the volume registered bythe PNT and that entering or leaving the lungs and hence adrift in volume This problem is most likely to occurimmediately after mask displacement if the infant movesor in a pressurized system (eg during artificialventilation)

Zero offset in flow calibration If the true zero flow isregistered as some finite value then integration of thisoffset over time results in a linear drift in volume with aslope equal to the offset Accurate delineation of the zeroflow point is more difficult as the sensitivity of the PNT

decreases which generally occurs as the linear rangeincreases The resolution of the AD converter used tosample the flow also sets a limit on how accurately the zeroflow point can be identified Therefore perfect offsetcompensation is never possible To prevent this volumedrift a dead band around the zero flow in which all valuesare set to zero is used in some devices However a deadband can hamper breath detection especially when flowis very low thus its use and the flow thresholds of thedead band should be described by the manufacturer of theequipment [21]

Imperfections in the pneumotachometer response If thetransducer for measuring flow does not function as aperfect measuring instrument (which is always the case tosome degree and may be significantly so under dynamicconditions) it is unlikely that the inspiratory and expiratoryflows are measured equally This produces asymmetries inthe recorded flow Such asymmetry can often be seen inmeasurements from infants intubated with small endo-tracheal tubes due to the geometric differences on eitherside of the PNT

Correcting volume drift

The analysis of tidal breathing data requires theexamination of data records containing a substantialnumber of breaths (typically $20) obtained during regularbreathing In principle it might be possible to avoid driftin volume in such a data record by preconditioning theinspired gas to BTPS conditions continuously monitoringgas partial pressures in both the alveoli and the pulmonaryarterial and venous blood to correct for respiratory ex-change ratios not equal to unity and eliminating all themethodological factors discussed above However this isextremely difficult if not impossible in practice Conse-quently it is never known how much of the baseline driftin volume is due to drift and how much represents a truechange in absolute lung volume Also because the subjectis assumed to be in the physiological steady state whendata are recorded the assumption is generally made thatfunctional residual capacity (FRC) remains more or lessconstant throughout the study period Such a situation isthus forced on the measured volume signal by some kindof drift correction algorithm which first assesses the driftand then removes it This does not of course mean thatFRC must be identical from one breath to the next butmerely that there is no net upward or downward trend inFRC over a period containing many breaths

Off-line drift correction algorithms commonly definethe drift in volume as the slope of the straight line fittedto the end-expiratory points in an epoch of tidal breaths(fig 3) In order to avoid any outliers skewing the reg-ression it may be useful to exclude those end-expiratorypoints with the greatest deviations and then refit the lineThis line is then subtracted from the volume to removethe drift and the mean level of the new end-expiratorypoints adjusted to zero There are also other ways inwhich volume can be drift-corrected such as subtractionof a curvilinear baseline instead of a straight line orrezeroing of volume at the end of every breath (whichrequires breath detection see below) Different dedrif-ting algorithms usually lead to slight differences in the

1185TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

for measurements in preterm infants should have a deadspace of lt15 mL Unfortunately the apparatus deadspace of most modern devices arises largely due to thenecessary connecting ports and so possibilities for deadspace reduction are limited In any case the connectionbetween PNT and mask must be minimized withoutdisturbing the linearity of the flowmeter [15] 2) Mini-mizing the resistance of the infant lung function equip-ment is important since the overall resistance of theequipment may not only dramatically change the respi-ratory pattern in spontaneously breathing babies butalso interfere with triggering devices in those who areventilated Any significant increase in resistance in-creases the expiratory time constant and potentiallyinfluences the end-expiratory level This in turn affectsany measurements that are volume-dependent includingvarious tidal breathing parameters The need to designfuture apparatus with as low a resistance as possiblewithin the constraints of simultaneously attaining a lowdead space and high resolution cannot be overempha-sized 3) The combined resistance of the apparatus (in-cluding any valves capnographs etc) should be lt20of the infantrsquos intrinsic resistance at the mean flowslikely to be encountered [1] Thus as a rough guide thecombined apparatus resistance should not exceed 12kPaL-1s at 50 mLs-1 in spontaneouslybreathing preterminfants 07 kPaL-1s at 100 mLs-1 in term neonates and05 kPaL-1s at 300 mLs-1 in infants and young children4) The response of the flowmeter should be linear overthe range of flows encountered The extent to which aPNT remains linear over an extended range is criticallydependent on design features such as whether it is acapillary- screen- or variable orifice-type device and thegeometry of any integral connections It is thereforeessential that the manufacturer provides accurate detailsand that the user checks the range of flows over whichthe flowmeter provides accurate recordings The approx-imate linearity ranges required for various sizes of infantare 0ndash100 mLs-1 in preterm infants and neonates of 2ndash4kg 0ndash300 mLs-1 in infants of 4ndash10 kg and 0ndash500 mLs-1

in preschool children of 11ndash15 kg In practice flowmeterswith a linear range of 0ndash10 Lmin-1 are commonly used inpreterm infants and neonates whereas those with a rangeof 0ndash35 Lmin-1 are used for obtaining tidal breathingmeasurements in older infants and young children 5) If aPNT has a nonlinear response over the desired flowrange it may be possible to effectively linearize it bycharacterizing the response and inverting it digitally Thiscan reduce dead space by allowing the use of a smaller-calibre PNT However it must be ensured that theresponse characteristics of the device remains constantover a prolonged period after repeated disinfection andon exposure to different respired gases 6) The flowmetershould have a flat frequency response over a frequencyrange sufficient to encompass the majority of the power inthe measured signals [1] For tidal breathing signals it isprobably sufficient to have a flat frequency response upto 10 Hz If the transducer itself does not have a flatfrequency response over this range it may be possibleto render it flat by digital compensation of the sampleddata [18] however this is only possible if the responseof the device is linear 7) If a PNT with metal screensor capillaries is used it should be heated to body tem-perature to avoid condensation on the resistive element

Major changes in screen resistance and hence measuredflows can occur within lt1 min of placing an unheatedPNT into a ventilator circuit thus this practice is stronglydiscouraged 8) Despite such heating PNTs with screensor capillaries in the ventilator circuit are highly suscep-tible to obstruction by secretions leading to falsely highmeasured flows and possible danger to the patientTherefore these PNTs should only be used by qualifiedpersonnel while the patient is under direct observation[19] 9) The geometry of the connectors on either side ofthe PNT screen affects the overall pressureflow charac-teristics of the device It is therefore important that theconnectors be as symmetrical as possible on either side ofthe PNT and that the PNT is calibrated in situ in exactlythe same configuration as that to be used with the subject[15] 10) If the inspired gas differs significantly fromroom air (eg by increased inspiratory oxygen fraction(FIO2)) it may be of different viscosity to room air andtherefore have different PNT calibration factors In such acase either the PNT should be calibrated with the inspiredgas or the room air calibration factors should be scaled bythe relative differences in gas viscosity [15] For measure-ments during artificial ventilation continuous FIO2correction at the bedside is advantageous [19] Theinfluence of changes in gas viscosity and density on thebehaviour of the PNT vary according to precise designand should be both stated by the manufacturer andchecked by the user

Data collection

Calibration of equipment

Equipment calibration has significant influence on thecalculated results and should be performed with the utmostcare and according to the recommendations of the manu-facturer Reliable measurements are unachievable with anunsuitable or defective calibration device It is thereforevital that 1) calibration is performed under identical cir-cumstances to and with the same equipment configurationas during measurements 2) the calibration tools arechecked periodically this requires that any calibratedsyringes or rotameters are returned to the manufacturers ofsuch devices on a regular basis according to the recom-mendations for any specific device (eg 12 monthly forprecision syringes) or whenever any deviation is suspec-ted 3) qualified personnel who understand both the pro-cedure and the signals and parameters displayed performthe calibration4) manual calibration is performed to checkthe automatic calibration procedures and 5) any deviationsin inspired gas viscosity are taken into account in the PNTcalibration

Data acquisition

Data acquisition requirements for infant respiratoryfunction testing are dealt with elsewhere in this series [3]Only those aspects of particular pertinence to tidalbreathing analysis are referred to below As discussedpreviously [3 15] it is crucial that the analogue flowsignal is passed through anti-aliasing filters withappropriate frequency cut-offs prior to sampling in order

1183TIDAL BREATH ANALYSIS IN INFANTS

to satisfy the Shannon sampling theorem and avoidthe potentially insidious problems of aliasing The flowdata are sampled by an analogue-to-digital (AD) con-verter which maps a specified voltage range in to anumber of equally spaced binary numbers It is crucialthat the incoming voltage signal from the flow transduceroccupies as much of the allowable voltage range ofthe AD converter as possible if maximum resolution is tobe attained For example if the flow ranges plusmn30 Lmin-1

and is digitized using a 12-bit AD converter themaximum resolution of the recorded flow signal is 60Lmin-1212 (ie 146 mLmin-1) For this reason togetherwith the need to minimize apparatus dead space andresistance a range of PNTs are probably needed toaccommodate infants of different ages undergoing dif-ferent types of respiratory function test in any one centreThe manufacturer should document both the flow rangeand number of bits of the AD converter

Sampling rate

The necessary sampling rate is determined by Shannonrsquostheorem and the clinical purpose of the tidal breathinganalysis The sampling interval (Dt) between flow datapoints determines the resolution of all identified timepoints such as the beginning and end of inspiration andexpiration Consequently identified time intervals suchas tI and tE have uncertainties of 2Dt For example with anfR of 60 breathsmin-1 and a sampling rate of 100 Hz(Dt=10 ms) the measurement error in tI and tE can be upto 4 A sampling rate of 100 Hz has been shown tobe normally adequate when calculating only VT and fR(see Appendix) whereas greater time resolution may berequired in rapidly breathing infants or for the measure-ment of certain parameters such as tPTEFtE Samplingrates of $200 Hz are therefore recommended foracquisition of tidal breathing data if such analyses are tobe performed particularly in small rapidly breathinginfants [8]

During data collection andor replay time-based dis-plays of flow and volume are required together withsimultaneous displays of flowvolume loops and relevanttrend data [20] These assist in the recognition of air leaksand behavioural state Of particular importance is thedetermination of when the infant has adapted to thepresence of the face mask Exactly how to make thisdetermination remains debatable The decision to com-mence recordings must be based on the operatorrsquos ownexperience plus observation of the displayed signals inorder to ascertain that 1) the breathing pattern is regularstable and representative for that infant 2) there is notrend in instantaneous fR (ie a stable mean fR has beenachieved) and 3) the signals are technically acceptable(eg no leaks artefacts or excessive volume drift)

Once the infant has adapted to the mask and is sleepingquietly and breathing regularly tidal breathing should berecorded in epochs of 30ndash60 s These should be repeatedover the next 5 min and at a later interval if a measure ofreproducibility is desired The number of recorded breat-hing cycles to use for evaluation depends on the variabilityof the signals but should allow the investigator to selectseveral epochs for evaluation It is recommended that eachepoch should contain at least 20 cycles

The essential general information which should berecorded when measuring any lung function parameters ininfants have been described previously [1] For a tidalbreathing study the additional data include time since lastfeeding start and end time of measurement andadaptation time

Signal processing

Numerical integration

Typically the primary measured signal is flow Thismust be integrated with respect to time to produce volumethis is most conveniently performed using a computerThere are a variety of numerical integration methodsavailable all of which connect adjacent data points withsome kind of curve and then sum the calculated areasbeneath each curve segment The more sophisticatedmethods make more accurate interpolations between thedata points than simpler algorithms but at the expense ofgreater complexity One of the simplest numerical inte-gration methods is the so-called trapezoidal rule (fig 2)This assumes that the sampled data points of the flowsignal are connected by straight lines and that the volumeincrement between the ith and (i-1)th data points (DVi) isgiven by

DVi = Dt (Vrsquoi + Vrsquoi 1)=2 (1)

where Vrsquoi and Vrsquoi-1 are the flow at the ith and (i-1)th datapoints The DVi are then summed to yield the total areaunder the curve

V = V0+Xn

i= 1

DVi (2)

where V0 is the volume at which integration of flow begins(which would normally be zero if integration begins at thestart of inspiration)

Flow

Time

Dt

V rsquoi-1V rsquoi

V rsquoi+1 V rsquoi+2

Fig 2 ndash Illustration of the trapezoidal rule The flow signal (mdash) isintegrated by joining its data points using straight lines (- - -) andcalculating the area under each line segment using Equation 1 volume increment between the ith and (i+1)th data points Dt samplinginterval Vrsquoi-1 Vrsquoi Vrsquoi+1 and Vrsquoi+2 flow at (i-1)th ith (i+1)th and (i+2)thdata points

1184 JHT BATES ET AL

Numerical integration is always in error when theoriginal continuous curve is represented by some kindof approximation function between the sampled pointsHowever these errors decrease as the data sampling rateincreases and the sampled points become more closelyspaced For most respiratory applications the integrationerror incurred with the trapezoidal rule is probablynegligible with a data sampling rate of 200 Hz

Volume drift

When flow is integrated to yield volume an upward ordownward drift in the volume baseline is invariably seenSome degree of drift is expected for purely physiologicalreasons For example the respiratory exchange ratio (iecarbon dioxide productionoxygen consumption)is usually~08 ie the volume of O2 absorbed by the lungs is 20greater than the volume of CO2 excreted This is reflectedin a slightly greater volume of gas being inspired thanexpired with each breath Also if the inspired air is notwarmed to body temperature and prehumidified thevolume of gas expired with each breath can be increasedby up to 5 (see discussion of body temperature baro-metric pressure and saturated with water vapour (BTPS)conditions below) relative to that inspired by a gain inwater vapour content These physiological effects con-tribute to a gradually increasing or decreasing volumemeasured at the mouth but not to a real change in baselinelung volume

In addition to the physiological factors discussed abovethe following methodological factors also contribute tovolume drift

Temperature changesbetween inspired and expired gas Ifinspired air is not warmed to body temperature beforepassing through the PNT it has a different viscosity anddensity to expired air which causes the PNT to registerthe transit of an equal number of molecules differentlybetween inspiration and expiration Variations in tem-perature may also affect the physical dimensions of thePNT due to the coefficients of thermal expansion of itscomponents

Changes in gas composition between inspiration andexpiration Inspired and expired gases differ in theirpartial pressures of O2 and CO2 This leads to slightdifferences in the viscosities of the gas mixtures withconcomitant effects on the flows registered duringinspiration and expiration by the PNT

Leaks Any leaks between the airway opening and PNTwhether through the mask seal or around a tracheal tubecause a discrepancy between the volume registered bythe PNT and that entering or leaving the lungs and hence adrift in volume This problem is most likely to occurimmediately after mask displacement if the infant movesor in a pressurized system (eg during artificialventilation)

Zero offset in flow calibration If the true zero flow isregistered as some finite value then integration of thisoffset over time results in a linear drift in volume with aslope equal to the offset Accurate delineation of the zeroflow point is more difficult as the sensitivity of the PNT

decreases which generally occurs as the linear rangeincreases The resolution of the AD converter used tosample the flow also sets a limit on how accurately the zeroflow point can be identified Therefore perfect offsetcompensation is never possible To prevent this volumedrift a dead band around the zero flow in which all valuesare set to zero is used in some devices However a deadband can hamper breath detection especially when flowis very low thus its use and the flow thresholds of thedead band should be described by the manufacturer of theequipment [21]

Imperfections in the pneumotachometer response If thetransducer for measuring flow does not function as aperfect measuring instrument (which is always the case tosome degree and may be significantly so under dynamicconditions) it is unlikely that the inspiratory and expiratoryflows are measured equally This produces asymmetries inthe recorded flow Such asymmetry can often be seen inmeasurements from infants intubated with small endo-tracheal tubes due to the geometric differences on eitherside of the PNT

Correcting volume drift

The analysis of tidal breathing data requires theexamination of data records containing a substantialnumber of breaths (typically $20) obtained during regularbreathing In principle it might be possible to avoid driftin volume in such a data record by preconditioning theinspired gas to BTPS conditions continuously monitoringgas partial pressures in both the alveoli and the pulmonaryarterial and venous blood to correct for respiratory ex-change ratios not equal to unity and eliminating all themethodological factors discussed above However this isextremely difficult if not impossible in practice Conse-quently it is never known how much of the baseline driftin volume is due to drift and how much represents a truechange in absolute lung volume Also because the subjectis assumed to be in the physiological steady state whendata are recorded the assumption is generally made thatfunctional residual capacity (FRC) remains more or lessconstant throughout the study period Such a situation isthus forced on the measured volume signal by some kindof drift correction algorithm which first assesses the driftand then removes it This does not of course mean thatFRC must be identical from one breath to the next butmerely that there is no net upward or downward trend inFRC over a period containing many breaths

Off-line drift correction algorithms commonly definethe drift in volume as the slope of the straight line fittedto the end-expiratory points in an epoch of tidal breaths(fig 3) In order to avoid any outliers skewing the reg-ression it may be useful to exclude those end-expiratorypoints with the greatest deviations and then refit the lineThis line is then subtracted from the volume to removethe drift and the mean level of the new end-expiratorypoints adjusted to zero There are also other ways inwhich volume can be drift-corrected such as subtractionof a curvilinear baseline instead of a straight line orrezeroing of volume at the end of every breath (whichrequires breath detection see below) Different dedrif-ting algorithms usually lead to slight differences in the

1185TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

to satisfy the Shannon sampling theorem and avoidthe potentially insidious problems of aliasing The flowdata are sampled by an analogue-to-digital (AD) con-verter which maps a specified voltage range in to anumber of equally spaced binary numbers It is crucialthat the incoming voltage signal from the flow transduceroccupies as much of the allowable voltage range ofthe AD converter as possible if maximum resolution is tobe attained For example if the flow ranges plusmn30 Lmin-1

and is digitized using a 12-bit AD converter themaximum resolution of the recorded flow signal is 60Lmin-1212 (ie 146 mLmin-1) For this reason togetherwith the need to minimize apparatus dead space andresistance a range of PNTs are probably needed toaccommodate infants of different ages undergoing dif-ferent types of respiratory function test in any one centreThe manufacturer should document both the flow rangeand number of bits of the AD converter

Sampling rate

The necessary sampling rate is determined by Shannonrsquostheorem and the clinical purpose of the tidal breathinganalysis The sampling interval (Dt) between flow datapoints determines the resolution of all identified timepoints such as the beginning and end of inspiration andexpiration Consequently identified time intervals suchas tI and tE have uncertainties of 2Dt For example with anfR of 60 breathsmin-1 and a sampling rate of 100 Hz(Dt=10 ms) the measurement error in tI and tE can be upto 4 A sampling rate of 100 Hz has been shown tobe normally adequate when calculating only VT and fR(see Appendix) whereas greater time resolution may berequired in rapidly breathing infants or for the measure-ment of certain parameters such as tPTEFtE Samplingrates of $200 Hz are therefore recommended foracquisition of tidal breathing data if such analyses are tobe performed particularly in small rapidly breathinginfants [8]

During data collection andor replay time-based dis-plays of flow and volume are required together withsimultaneous displays of flowvolume loops and relevanttrend data [20] These assist in the recognition of air leaksand behavioural state Of particular importance is thedetermination of when the infant has adapted to thepresence of the face mask Exactly how to make thisdetermination remains debatable The decision to com-mence recordings must be based on the operatorrsquos ownexperience plus observation of the displayed signals inorder to ascertain that 1) the breathing pattern is regularstable and representative for that infant 2) there is notrend in instantaneous fR (ie a stable mean fR has beenachieved) and 3) the signals are technically acceptable(eg no leaks artefacts or excessive volume drift)

Once the infant has adapted to the mask and is sleepingquietly and breathing regularly tidal breathing should berecorded in epochs of 30ndash60 s These should be repeatedover the next 5 min and at a later interval if a measure ofreproducibility is desired The number of recorded breat-hing cycles to use for evaluation depends on the variabilityof the signals but should allow the investigator to selectseveral epochs for evaluation It is recommended that eachepoch should contain at least 20 cycles

The essential general information which should berecorded when measuring any lung function parameters ininfants have been described previously [1] For a tidalbreathing study the additional data include time since lastfeeding start and end time of measurement andadaptation time

Signal processing

Numerical integration

Typically the primary measured signal is flow Thismust be integrated with respect to time to produce volumethis is most conveniently performed using a computerThere are a variety of numerical integration methodsavailable all of which connect adjacent data points withsome kind of curve and then sum the calculated areasbeneath each curve segment The more sophisticatedmethods make more accurate interpolations between thedata points than simpler algorithms but at the expense ofgreater complexity One of the simplest numerical inte-gration methods is the so-called trapezoidal rule (fig 2)This assumes that the sampled data points of the flowsignal are connected by straight lines and that the volumeincrement between the ith and (i-1)th data points (DVi) isgiven by

DVi = Dt (Vrsquoi + Vrsquoi 1)=2 (1)

where Vrsquoi and Vrsquoi-1 are the flow at the ith and (i-1)th datapoints The DVi are then summed to yield the total areaunder the curve

V = V0+Xn

i= 1

DVi (2)

where V0 is the volume at which integration of flow begins(which would normally be zero if integration begins at thestart of inspiration)

Flow

Time

Dt

V rsquoi-1V rsquoi

V rsquoi+1 V rsquoi+2

Fig 2 ndash Illustration of the trapezoidal rule The flow signal (mdash) isintegrated by joining its data points using straight lines (- - -) andcalculating the area under each line segment using Equation 1 volume increment between the ith and (i+1)th data points Dt samplinginterval Vrsquoi-1 Vrsquoi Vrsquoi+1 and Vrsquoi+2 flow at (i-1)th ith (i+1)th and (i+2)thdata points

1184 JHT BATES ET AL

Numerical integration is always in error when theoriginal continuous curve is represented by some kindof approximation function between the sampled pointsHowever these errors decrease as the data sampling rateincreases and the sampled points become more closelyspaced For most respiratory applications the integrationerror incurred with the trapezoidal rule is probablynegligible with a data sampling rate of 200 Hz

Volume drift

When flow is integrated to yield volume an upward ordownward drift in the volume baseline is invariably seenSome degree of drift is expected for purely physiologicalreasons For example the respiratory exchange ratio (iecarbon dioxide productionoxygen consumption)is usually~08 ie the volume of O2 absorbed by the lungs is 20greater than the volume of CO2 excreted This is reflectedin a slightly greater volume of gas being inspired thanexpired with each breath Also if the inspired air is notwarmed to body temperature and prehumidified thevolume of gas expired with each breath can be increasedby up to 5 (see discussion of body temperature baro-metric pressure and saturated with water vapour (BTPS)conditions below) relative to that inspired by a gain inwater vapour content These physiological effects con-tribute to a gradually increasing or decreasing volumemeasured at the mouth but not to a real change in baselinelung volume

In addition to the physiological factors discussed abovethe following methodological factors also contribute tovolume drift

Temperature changesbetween inspired and expired gas Ifinspired air is not warmed to body temperature beforepassing through the PNT it has a different viscosity anddensity to expired air which causes the PNT to registerthe transit of an equal number of molecules differentlybetween inspiration and expiration Variations in tem-perature may also affect the physical dimensions of thePNT due to the coefficients of thermal expansion of itscomponents

Changes in gas composition between inspiration andexpiration Inspired and expired gases differ in theirpartial pressures of O2 and CO2 This leads to slightdifferences in the viscosities of the gas mixtures withconcomitant effects on the flows registered duringinspiration and expiration by the PNT

Leaks Any leaks between the airway opening and PNTwhether through the mask seal or around a tracheal tubecause a discrepancy between the volume registered bythe PNT and that entering or leaving the lungs and hence adrift in volume This problem is most likely to occurimmediately after mask displacement if the infant movesor in a pressurized system (eg during artificialventilation)

Zero offset in flow calibration If the true zero flow isregistered as some finite value then integration of thisoffset over time results in a linear drift in volume with aslope equal to the offset Accurate delineation of the zeroflow point is more difficult as the sensitivity of the PNT

decreases which generally occurs as the linear rangeincreases The resolution of the AD converter used tosample the flow also sets a limit on how accurately the zeroflow point can be identified Therefore perfect offsetcompensation is never possible To prevent this volumedrift a dead band around the zero flow in which all valuesare set to zero is used in some devices However a deadband can hamper breath detection especially when flowis very low thus its use and the flow thresholds of thedead band should be described by the manufacturer of theequipment [21]

Imperfections in the pneumotachometer response If thetransducer for measuring flow does not function as aperfect measuring instrument (which is always the case tosome degree and may be significantly so under dynamicconditions) it is unlikely that the inspiratory and expiratoryflows are measured equally This produces asymmetries inthe recorded flow Such asymmetry can often be seen inmeasurements from infants intubated with small endo-tracheal tubes due to the geometric differences on eitherside of the PNT

Correcting volume drift

The analysis of tidal breathing data requires theexamination of data records containing a substantialnumber of breaths (typically $20) obtained during regularbreathing In principle it might be possible to avoid driftin volume in such a data record by preconditioning theinspired gas to BTPS conditions continuously monitoringgas partial pressures in both the alveoli and the pulmonaryarterial and venous blood to correct for respiratory ex-change ratios not equal to unity and eliminating all themethodological factors discussed above However this isextremely difficult if not impossible in practice Conse-quently it is never known how much of the baseline driftin volume is due to drift and how much represents a truechange in absolute lung volume Also because the subjectis assumed to be in the physiological steady state whendata are recorded the assumption is generally made thatfunctional residual capacity (FRC) remains more or lessconstant throughout the study period Such a situation isthus forced on the measured volume signal by some kindof drift correction algorithm which first assesses the driftand then removes it This does not of course mean thatFRC must be identical from one breath to the next butmerely that there is no net upward or downward trend inFRC over a period containing many breaths

Off-line drift correction algorithms commonly definethe drift in volume as the slope of the straight line fittedto the end-expiratory points in an epoch of tidal breaths(fig 3) In order to avoid any outliers skewing the reg-ression it may be useful to exclude those end-expiratorypoints with the greatest deviations and then refit the lineThis line is then subtracted from the volume to removethe drift and the mean level of the new end-expiratorypoints adjusted to zero There are also other ways inwhich volume can be drift-corrected such as subtractionof a curvilinear baseline instead of a straight line orrezeroing of volume at the end of every breath (whichrequires breath detection see below) Different dedrif-ting algorithms usually lead to slight differences in the

1185TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

Numerical integration is always in error when theoriginal continuous curve is represented by some kindof approximation function between the sampled pointsHowever these errors decrease as the data sampling rateincreases and the sampled points become more closelyspaced For most respiratory applications the integrationerror incurred with the trapezoidal rule is probablynegligible with a data sampling rate of 200 Hz

Volume drift

When flow is integrated to yield volume an upward ordownward drift in the volume baseline is invariably seenSome degree of drift is expected for purely physiologicalreasons For example the respiratory exchange ratio (iecarbon dioxide productionoxygen consumption)is usually~08 ie the volume of O2 absorbed by the lungs is 20greater than the volume of CO2 excreted This is reflectedin a slightly greater volume of gas being inspired thanexpired with each breath Also if the inspired air is notwarmed to body temperature and prehumidified thevolume of gas expired with each breath can be increasedby up to 5 (see discussion of body temperature baro-metric pressure and saturated with water vapour (BTPS)conditions below) relative to that inspired by a gain inwater vapour content These physiological effects con-tribute to a gradually increasing or decreasing volumemeasured at the mouth but not to a real change in baselinelung volume

In addition to the physiological factors discussed abovethe following methodological factors also contribute tovolume drift

Temperature changesbetween inspired and expired gas Ifinspired air is not warmed to body temperature beforepassing through the PNT it has a different viscosity anddensity to expired air which causes the PNT to registerthe transit of an equal number of molecules differentlybetween inspiration and expiration Variations in tem-perature may also affect the physical dimensions of thePNT due to the coefficients of thermal expansion of itscomponents

Changes in gas composition between inspiration andexpiration Inspired and expired gases differ in theirpartial pressures of O2 and CO2 This leads to slightdifferences in the viscosities of the gas mixtures withconcomitant effects on the flows registered duringinspiration and expiration by the PNT

Leaks Any leaks between the airway opening and PNTwhether through the mask seal or around a tracheal tubecause a discrepancy between the volume registered bythe PNT and that entering or leaving the lungs and hence adrift in volume This problem is most likely to occurimmediately after mask displacement if the infant movesor in a pressurized system (eg during artificialventilation)

Zero offset in flow calibration If the true zero flow isregistered as some finite value then integration of thisoffset over time results in a linear drift in volume with aslope equal to the offset Accurate delineation of the zeroflow point is more difficult as the sensitivity of the PNT

decreases which generally occurs as the linear rangeincreases The resolution of the AD converter used tosample the flow also sets a limit on how accurately the zeroflow point can be identified Therefore perfect offsetcompensation is never possible To prevent this volumedrift a dead band around the zero flow in which all valuesare set to zero is used in some devices However a deadband can hamper breath detection especially when flowis very low thus its use and the flow thresholds of thedead band should be described by the manufacturer of theequipment [21]

Imperfections in the pneumotachometer response If thetransducer for measuring flow does not function as aperfect measuring instrument (which is always the case tosome degree and may be significantly so under dynamicconditions) it is unlikely that the inspiratory and expiratoryflows are measured equally This produces asymmetries inthe recorded flow Such asymmetry can often be seen inmeasurements from infants intubated with small endo-tracheal tubes due to the geometric differences on eitherside of the PNT

Correcting volume drift

The analysis of tidal breathing data requires theexamination of data records containing a substantialnumber of breaths (typically $20) obtained during regularbreathing In principle it might be possible to avoid driftin volume in such a data record by preconditioning theinspired gas to BTPS conditions continuously monitoringgas partial pressures in both the alveoli and the pulmonaryarterial and venous blood to correct for respiratory ex-change ratios not equal to unity and eliminating all themethodological factors discussed above However this isextremely difficult if not impossible in practice Conse-quently it is never known how much of the baseline driftin volume is due to drift and how much represents a truechange in absolute lung volume Also because the subjectis assumed to be in the physiological steady state whendata are recorded the assumption is generally made thatfunctional residual capacity (FRC) remains more or lessconstant throughout the study period Such a situation isthus forced on the measured volume signal by some kindof drift correction algorithm which first assesses the driftand then removes it This does not of course mean thatFRC must be identical from one breath to the next butmerely that there is no net upward or downward trend inFRC over a period containing many breaths

Off-line drift correction algorithms commonly definethe drift in volume as the slope of the straight line fittedto the end-expiratory points in an epoch of tidal breaths(fig 3) In order to avoid any outliers skewing the reg-ression it may be useful to exclude those end-expiratorypoints with the greatest deviations and then refit the lineThis line is then subtracted from the volume to removethe drift and the mean level of the new end-expiratorypoints adjusted to zero There are also other ways inwhich volume can be drift-corrected such as subtractionof a curvilinear baseline instead of a straight line orrezeroing of volume at the end of every breath (whichrequires breath detection see below) Different dedrif-ting algorithms usually lead to slight differences in the

1185TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

subsequently estimated values of breathing pattern para-meters However given that drift correction is merely anempirical operation it is probably appropriate to select acorrection algorithm on the basis of robustness and easeof implementation as much as anything else For the userof breath analysis software it is important to know whichmethod of drift correction has been implemented It isalso useful to be able to switch off the correction pro-cedure so that real changes in FRC can be tracked overshort periods when the recorded flow signal is sufficientlyaccurate

A particular problem with automated drift correctionalgorithms is that they can mask the presence of significantdifferences between inspiratory and expiratory VT such asmight arise from air leaks in the breathing circuit or severePNT asymmetries Consequently the magnitude of thecorrection made for volume drift should always be moni-tored The drift magnitude (Drift) is usefully defined as themean drift per breath divided by the mean VT over theepoch of volume being analysed given as a percentage by

Drift = 100DV

PN

i= 1VTi

(3)

where DV is the drift of the volume baseline over Ncomplete breaths and VTi is the ith VT An unusually largedrift magnitude is indicative of a methodological problemsuch as the presence of a large air leak

The end-expiratory lung volume has a significantinfluence on many tidal breathing parameters [2 4 5]Once the volume has been dedrifted zero volume isgenerally defined as the mean end-expiratory level (EEL)This should be displayed on the time-based trace toensure that it is representative of the data with the userbeing given the option to adjust it if necessary

The variability of the end-expiratory values that arescattered around the zero EEL can then be used to assessthe stability of the EEL A preliminary suggestion is thatafter correcting the volume signal for drift the mean EELis established from 20ndash30 breaths with the variation in

individualend-expiratorypoints from this mean EEL beingused to calculate the SD of the EEL This could then beexpressed relative to the absolute magnitude of the VT forwithin- and between-subject comparisons

During on-line measurements rezeroing of the volumeat the beginning of inspiration may be helpful in stabilizingthe display However during off-line evaluation the trueEEL after drift correction should be used so that breath-to-breath variations can be detected The practice ofpresenting inspired and expired volumes separately (ieabove and below the zero axis respectively) should bediscouraged since much information about the breathingpattern is lost including any instability of the EEL orvolume drift Whichever procedure is used to stabilize thevolume signal for display there should always be themeans to disable this and the user must ensure that theprocedure does not mask the presence of leaks

Body temperature barometric pressure and saturatedwith water vapour conditions

Errors of up to 11 may occur if inspiratory flow andvolume are not converted to BTPS conditions If VT isexpressed as the mean of the inspired (VTI) and expiredvolumes (VTE) which is the recommended practice exceptin intubated babies in whom inspiratory leak may be aproblem neglecting to convert this to BTPS conditionsleads to an underestimation of ~5 Unfortunately cor-recting to BTPS conditions is not always straightforwardFor example if tidal breathing measurements are madewhen backgroundgas flow is superimposed on the exhaledgas the precise BTPS correction factor to apply may beunknown [22] Also although it is generally assumed thatexpired gas is at BTPS conditions there may be somedeconditioning of expired gas before it reaches the PNTin practice (personal communication J ReinstaedtlerInternational Applications Erich Jaeger GmbH Hoch-berg Germany) For the purposes of standardization andto avoid any systematic bias between different systemsfor assessing tidal breathing and other parameters ofrespiratory function in infants it is currently recom-mended that inspiratory flow be corrected to BTPSconditions using the following equation

VrsquoBTPS = VrsquoATP Tb (P amb

P rsquoH2Oamb)=Tamb (P amb PH2OTb) (4)

where VrsquoBTPS and VrsquoATP are flow under BTPS and ambienttemperature and barometric pressure conditions Tb andTamb are the thermodynamic body (3102 K) and ambienttemperature and Pamb PH2OTb and PH2Oamb are theambient pressure and water the vapour pressure at 100humidity at Tb (63 kPa) and of the ambient gas the lattercan be approximated by

P H2Oamb = (RHamb P H2Ox)=100 (5)

where RHamb is the relative humidity of the ambient gas(as a percentage) and PH2Ox the water vapour pressure at100 humidity at a temperature of x

0 2 4 6 8 10 12 14 16Time s

-505

1015202530354045

Volu

me

mL

Fig 3 ndash Example of volume drift correction The oscillating volumesignal drifting upwards (mdash) has a straight line that characterizes its driftSubtracting this line from the volume yields a drift-corrected signal thatoscillates about a stable baseline (- - -)

1186 JHT BATES ET AL

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

BTPS corrections are obviously not necessary when theinfant inspires air which has been preconditioned to BTPSconditions such as during plethysmographic measure-ments of airway resistance using a heated rebreathing bag[5] However data collected under the latter conditionsare unsuitable for tidal breathing analysis due to theinevitable stimulation of breathing under these conditions[23]

Automatic breath identification

One of the main challenges in tidal breathing analysisfrom the point of view of the computer programmer isthe automatic identification of the beginning of inspi-ration and expiration for each breath in a series Thisinvolves pattern recognition which is notoriously difficultfor computers even though human observers might findthe task easy Indeed it is no trivial matter to come upwith an algorithm that works all the time and never missesbreaths or identifies ones that do not exist Once theindividual inspirations and expirations have been identi-fied determining VT tI tE and fR for each breath isessentially straightforward

In recent years there have been several attempts toidentify the most robust type of breath identificationalgorithm [24] The most frequently used algorithm forbreath detection is based on flow thresholds as shown infigure 4 The choice of flow threshold is critical because itmust be higher than the noise level in order to preventfalse triggering but low enough to detect small breathsIdeally the flow threshold should depend on age butmost algorithms use a fixed value for all subjects Thiscan lead to poor identification of breaths especially insmall babies with rapid or irregular breathing patterns Incommercial devices the flow threshold used should beclearly given by the manufacturer together with anyplausibility tests [24] Once the flow threshold has beencrossed a threshold algorithm must search back to the lastzero crossing of flow to find the precise time of theinspiratoryexpiratory transition [7 8]

The Appendix analyses two breath detection algorithmsthat have been used in a number of previous investigationsOne algorithm identifies the zero crossings of a smoothedflow signal and the other identifies zero crossings in flows

that bracket peak flow magnitudes above a set thresholdAlthough these algorithms do not represent everything thatis possible in automatic breath detection they serve toillustrate some of the key problems involved and de-monstrate that different algorithms can perform differentlyunder certain circumstances

Data evaluation and reporting

Tidal breathing measurements should be accompaniedby high-resolution graphic display showing the measuredflow and volume signals plotted against time and againsteach other These plots should be of sufficient clarity toallow manual validation of the calculated breathing pat-tern parameters because despite apparently clear defini-tions correct measurement of these parameters is often notstraightforward In particular automatic determination ofthe start and end of each breath can be unreliable if theshape of the volume signal differs significantly from thetypical normal form shown in figure 1b (see Appendix)Automatic breath detection should thus be accompaniedby visual confirmation which requires adequate display ofthe measured signals something which has rarely beenavailable from commercial devices in the past

Evaluation of the measured flow and volume signals iscommonly performed off-line beginning with replay ofthe stored signals and selection by the operator of suitableepochs for analysis Data that are not accepted for ana-lysis should not be deleted as they may be valuable inretrospect In the final report of a tidal breathing analysisthe total number of breaths recorded and the number ofthese selected for analysis should be given The mean andSD or coefficient of variation should be reported for allparameters The report should also include essential patientcharacteristics [1] representative time-based signals andflowvolume loops together with a parameter table ofindividual trials and a statistical summary

Reference data

In order to use tidal breathing analysis effectively in theclinical setting it is important to know 1) the influence ofgrowth and maturation (including gestational and post-natal age) on the various tidal breathing parameters 2)the influence of demographic factors such as sex andethnic group on tidal breathing parameters 3) the normalintra- and interindividual variability of the parameters atevery age and 4) the diagnostic value (if any) of thevarious parameters Unfortunately despite repeated effortsover the last 50 years to establish reference values forventilatory parameters in healthy infants knowledgeregarding the biological development and clinicaldiag-nostic value of most tidal breathing parameters remainssparse Although some so-called reference data havebeen published these values are highly specific to theequipment used and the behavioural state of the specificpopulation studied and cannot be recommended for gene-ral use This problem needs to be addressed urgently in thenear future once equipment and measurement conditionshave been standardized

Time

Insp

Exp

Flow 0

d

Fig 4 ndash Illustration of a flow threshold (- - -) algorithm for thedetection of the beginningof inspiration (insp) and expiration (exp) recognized start of inspexp x verification of inspexp d time delaybetween recognition and verification

1187TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

Conclusions

The study of tidal breathing in infants and childrenbegins with the measurement of flow at the mouth duringquiet breathing When collecting flow data it is importantthat 1) a snugly fitting face mask is used to minimize airleaks with the dead space of the mask being estimatedat 50 of its physical volume 2) a flowmeter withappropriate frequency response and linearity is employed3) efforts are made to eliminate the various sources ofdrift in volume that occur when flow is integrated withrespect to time 4) the AD converter used to samplethe flow signal can adequately resolve the largest andsmallest flows required by the study 5) the flow is filteredfor anti-aliasing and sampled so as to satisfy the samplingtheorem (a sampling rate of 100 Hz appears to be sufficientfor the determination of time and volume parameters butrates of 200 Hz are necessary for analysis of the tidalbreathing flowvolume loop and other sensitive parameterssuch as tPTEFtE) and 6) inspired gases are corrected toBTPS conditions

Once the data have been collected certain key signalprocessing considerations must be attended to as follows1) flow must be integrated to obtain volume using anappropriate numerical integration algorithm (trapezoidalintegration is sufficiently accurate for most applicationsinvolving data sampled at $100 Hz) 2) a drift correctionalgorithm must be employed to eliminate the inevitableupward or downward drift of the volume baselineobtained from integrating flow and 3) the magnitude ofthe drift in volume should be monitored for indications ofa possible air leak

The potentially most troublesome aspect of tidal breathanalysis from the computational point of view is theidentification of the beginning and end of inspiration andexpiration It would clearly be advantageous if the com-puter were to perform this labour-intensive task Howeverit may not be possible to devise a completely automaticalgorithm that works satisfactorily in every case thus somemeans of quality control by visual inspection is desirable toensure appropriate segmentation of individual breathsOnce the individual breaths in a flow record have beensuccessfully identified it is relatively straightforward tocalculate the various indices of the breathing pattern thatmay be of interest

Once the methods and equipment for measuring andanalysing tidal breathing in infants are standardized thereis an urgent need to establish appropriate reference rangesfor various key parameters so that they may be used moreeffectively in the clinical setting

Appendix automatic breath identification

In this appendix two algorithms for automatic breathidentification are examined in order to illustrate some ofthe issues and difficulties involved These algorithms are1) an algorithm that identifies the zero crossings of asmoothed flow signal the smoothed algorithm and 2)an algorithm that identifies zero crossings in flow thatbracket peak flow magnitudes above a set threshold thethreshold algorithm

The smoothed algorithm

This algorithm identifies the beginning of inspirationand expiration in each breath from the points at which flowcrosses zero This requires two conditions to be satisfied1) there is no significant zero offset in flow and 2) flowonly changes sign at the inspiratoryexpiratory transitionsThe first condition is ensured with reasonable accuracy ina first pass over a data record containing a number ofbreaths by subtracting the mean of the flow signal fromitself The second condition is more problematic becausecardiogenic oscillations in flow together with other extra-neous noise sources can cause flow to cross zero atmultiple points within a breath This is particularly pre-valent at the end of expiration at which point themagnitude of flow is low For this reason the smoothedalgorithm first identifies the beginning of expiration ineach breath corresponding to the peaks in volume as theseare generally less obscured by extraneous oscillations

Breath identificationTo eliminate the problem of spuriouszero crossings the flow signal is first smoothed bycalculating its running mean using a window length of Ndata points The smoothed signal (Vrsquos) is then

Vrsquos(i) = (1=N)Xi+ N=2

j= i N=2

Vrsquoj (6)

For N=246 the smoothed flow signal tends to showfewer high-frequency noise-generated oscillations than theflow signal If N is chosen properly only the low-frequency oscillations in flow corresponding to completebreaths are left in the smoothed flow Figure 5 showsan example of the result of this smoothing operation onthe flow signal from a single breath Figure 5 also showsthat the smoothing operation shifts the positions of thezero crossings Therefore the crossings in the smoothedflow signal cannot be taken as the final positions ofthe inspiratoryexpiratory transitions in flow The finalpositions of the beginning of expiration are found asfollows First the smoothed flow signal is examined for

-1000

0

500

1000

0 1 2 3 4 5Time s

Flow

mLmiddots

-1

Fig 5 ndash Smoothing a signal using a running mean Note the multiplezero crossings towards the end of expiration in the original flow signal(mdash) The smoothed flow signal (- - -) was obtained by smoothing flowusing a 1-s running mean eliminating the multiple zero crossings

1188 JHT BATES ET AL

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

those points at which it crosses zero from positive tonegative These points are close to but generally notcoincident with the beginning of each expirationTherefore the flow signal is next integrated to obtainvolume and the positions of the volume minima arelocated between those time points at which the begin-nings of expiration were previously located in thesmoothed flow signal The regions between these volumeminima are then researched for their maxima which arethe true beginnings of expiration

Note that if N is too small not all the spurious zerocrossings will be eliminated in the smoothed flow signalby the above procedure Similarly if N is too large someof the real breaths may be eliminated

Generally speaking N should correspond to a windowlength of approximately one breath period but thisobviously varies with the particular data record beinganalysed It is therefore not possible to specify a singlevalue of N that works in every case For this reason thesmoothed algorithm interacts with the operator for thedetermination of N The operator is prompted for a suit-able value and is then shown the resulting breathidentification as a volume signal If the operator decidesthat some breaths have been missed or incorrectlyidentified a different value of N may be tried This pro-cess is repeated until breath identification is satisfactoryThis interactive process may not be suitable for generalclinical applications as the operator may not have theexpertise andor the time necessary to go through thevisual quality check procedure described above Forgeneral use it may therefore be best to use a defaultlength for the smoothing window that works well in mostsituations Nevertheless it is clearly advisable to havesome means of resorting to visual quality control so thatan expert can deal with questionable cases that have notbeen dealt with unambiguously by the algorithm

Volume drift correction At this stage however thevolume signal will probably still contain some residualdrift that has failed to be eliminated by subtraction ofthe mean flow This residual drift is removed by identifyingthe volume minima between each inspiratoryexpiratorytransition and then adding a constant to the flow so thatwhen it is reintegrated these volume minima lie along aregression line with a slope of zero In other words theFRC is forced to vary about a horizontal baseline (fig 3)Finally the maxima and minima for each breath areidentified in the drift-corrected volume signal

At this point the volume signal has been corrected fordrift and the beginning of each expiration identifiedtogether with the volume minima for each breath TheVT for each breath is then simply the mean of VTI (thedifference in volume between the beginning of the cor-responding expiration and the preceding lowest point)and VTE (the difference in volume between the begin-ning of expiration and the subsequent lowest point)Identifying tI and tE might seem equally trivial merelyrequiring identification of the time differences betweeneach inspiratoryexpiratory transition and its preceding orsucceeding lowest point However even with smoothingcardiogenic oscillations in the tail of a long slow expira-tion can produce significant variation in the timing of thelowest point because the volume signal is so flat in thisregion For this reason tI is determined by starting at each

transition from inspiration to expiration and workingbackwards until the volume comes within 5 of thelowest point in the preceding breath The time intervalbetween this point and the start of the preceding expirationis taken as tE

The threshold algorithm

Breath identification The threshold and smoothedalgorithms both identify transitions between inspirationand expiration from zero crossings in flow Howeverspurious crossings such as those due to cardiogenicoscillations toward the end of expiration are eliminated ina different manner in the threshold algorithm In thisalgorithm all zero crossings in flow are first identifiedNext the peak magnitudes of flow either positive ornegative between each zero crossing are found Finallypairs of zero crossings are discarded if they are separatedby a peak flow whose magnitude is less than a certainthresholdFigure 6 shows the zero crossings in flow from asingle breath The crossings separated by low-magnitudepeaks are discarded The flow threshold for discardingzero crossings varies according to age In preterm infantsand newborns it may be as low as 10 mLs-1 whereas ininfants beyond the neonatal period (gt4 kg) a threshold of~30 mLs-1 usually works well As with the smoothedalgorithm a fixed threshold may not work in every casethus the user is given the option of changing the thresholduntil satisfactory results are obtained Unfortunately it isnot possible to specify a single threshold that will suitevery situation If the threshold is too low false breathsmay be detected but if the threshold is too high realbreaths may be missed The most robust algorithms arethose in which a flow threshold is combined with someadditional plausibility criteria [24]

Volume drift correction Volume drift correction isachieved by the threshold algorithm in the same way asby the smoothed algorithm that is the end-expiratoryvolumes are made to lie along a horizontal regression lineVT is obtained identically by both algorithms Using thethreshold algorithm tI and tE are determined from thetime intervals between successive zero crossings which

0 5Time s

0Flow

Fig 6 ndash Zero crossings in flow (vertical lines) Those crossings sepa-rated by low-magnitude peaks in flow (in this case those toward the endof expiration) are discarded leaving only those crossings that define thetransitions between inspiration and expiration

1189TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

differs somewhat from the smoothed algorithm Note thatthe time resolution of both algorithms as used in thepresent study are determined by the data sampling ratebecause zero crossings in flow are determined to thenearest data point With a data sampling rate of 100 Hz forexample going to the nearest data point gives timingestimates accurate to within 10 ms which is probablysufficient when simply measuring VT and fR but couldintroduce significant errors when determining short tPTEFSince greater accuracy in zero crossing determination iseasily obtained by interpolating between the two datapoints that span zero this should probably be carried out asa general rule

Comparison of smoothed and threshold breathidentification algorithms

Figure 7 shows two 40-s records of flow used to test thealgorithms These records were obtained from infants of34ndash38 weeks gestational age and represent two typesof signal 1) regular ventilation in which the individualbreaths are clear and well defined and 2) ventilation inwhich there are large rapid oscillations in flow within oneof the breaths towards the end of the record (this wasdeliberately chosen as an extreme example for testingthe algorithms and would not be considered a suitableepochs for the analysis of tidal breathing parameters whenused to reflect lung mechanical properties)

Neither the smoothed nor the threshold algorithmpresented any difficulty in correctly identifying the breathsshown in figure 7a The VT tI and tE obtained using thetwo algorithms are given in table 1 The slight differencesbetween the values of some of the parameters returnedby the two algorithms are no doubt due to the differencesin the way that the beginning of inspiration is defined Inthe smoothed algorithm it is the point at which thevolume comes within 5 of its lowest point when work-ing backwards from the peak whereas in the thresholdalgorithm it is simply the lowest point in volume Thusthe threshold algorithm gives a more accurate estimatewhen the troughs in volume are well defined as in thedata set considered The smoothed algorithm in contrastis more robust to the presence of cardiogenic oscillationsat the end of a long expiration when flow is low andvolume is sensitive to having its minimum displaced alarge distance by noise

The two algorithms did not fare equally whenconsidering the flow record shown in figure 7b howeverThe smoothed algorithm easily identified the breathscorrectly but the large rapid oscillations in flow towardsthe end of the record caused problems for the thresholdalgorithm These oscillations do not correspond to truebreaths as figure 8 clearly shows However the thresholdalgorithm was unable to eliminate them as candidatesbecause their peak magnitudes were comparable to thoseof real breaths and so they were not detected by the flowthreshold This example illustrates the key differencebetween the way in which the smoothed and thresholdalgorithms operate The smoothed algorithm involves afiltering operation that manipulates the frequency contentof flow in order to separate spurious high-frequencyevents from lower-frequency true breaths The thresholdalgorithm in contrast considers the amplitude character-istics of flow which in this example are similar for bothspurious oscillations and true breaths

It thus appears that the threshold algorithm is unable tofunction successfully in all cases in which the smoothedalgorithm does succeed indicating that using a frequencyfiltering operation to identify breaths is better than using anapproach based simply on amplitude discrimination incases in which periods of irregular breathing are to beexamined This is particularly relevant in cases in whichlong-term recordings to investigate regulation of breathingpatterns are being undertaken During routine respiratoryfunction testing this should be less of a problem as theoperator should select epochs of regular breathing

-150

0

150

Flow

mLmiddots

-1

a)

-300

0

300

Flow

mLmiddots

-1

b)

0 40Time s

Fig 7 ndash The two flow records used to test the breath identificationalgorithms a) clear well-defined breaths and b) somewhat less regularbreaths with in particular some high-amplitude rapid oscillations in thethird-from-last breath

Table 1 ndash Tidal breathing parameters identified from 18consecutive breaths determined by the smoothed andthreshold algorithms

Smoothed Threshold

VT mL 666plusmn37 666plusmn36tI s 075plusmn007 076plusmn007tE s 131plusmn019 129plusmn019

Data are presented as meanplusmnSD from flow record shown infigure 7a The smoothed algorithm was implemented using a 2-window to calculate the running mean the threshold algorithm wasimplemented with a flow threshold of 10 mLs-1 VT tidal volumetI inspiratory time tE expiratory time

1190 JHT BATES ET AL

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

SCHMIDT et al [24] investigated a number of algorithmsapplied to newborns and found similar results in mostcases all algorithms agreed but there were always a fewdifficult situations in which some algorithms outper-formed others Thus it seems somewhat doubtful analgorithm can ever be devised that works all the timewithout operator intervention Nevertheless it wouldclearly be advantageous to have an algorithm for clinicaluse that functions at least most of the time without userinteraction One possibility for dealing with difficultcases would be to use the combination of a thresholdalgorithm with various feasibility criteria [24] and todiscard all identified breaths that fall outside some agreedrange (eg plusmn10) of the mean VT or ttot

Effects of data sampling rate

The rate at which a flow or volume signal is sampledinfluences the values of breathing pattern parameters Alow sampling rate obviously gives reduced temporalresolution of timing parameters such as tI and tE as shownabove In order to test the influence of data sampling rateon estimation of VT tI tE and ttot the smoothed algorithmwas used to analyse the flow records shown in figure 7when resampled at 75 50 and 25 Hz (the original

sampling rate being 100 Hz) In both cases the flowsignals were analysed using a smoothing window of 2 s

VT was particularly insensitive to changes in datasampling rate as might be expected because it is a measureof signal amplitude rather than timing The mean VTobtained from each of the test signals did not changeby more than 01 as the sampling rate was dropped from100 to 25 Hz The timing parameters tI and tE weresomewhat more affected changing by up to 07 as thesampling rate dropped from 100 to 75 Hz by up to 2 asthe rate dropped to 50 Hz and by up to 35 as the ratedropped to 25 Hz Interestingly the remaining timingparameter ttot changed by only up to 01 over this rangeof sampling rates indicating that errors in estimating tIwere compensated for by virtually equal and oppositeerrors in tE These results suggest that a data samplingrate of 100 Hz is adequate for accurate estimation of VT tIand tE (and hence fR) Indeed for most applications arate of 50 Hz is probably adequate if it is only theseparameters that are to be analysed Nevertheless whentiming parameters such as tPTEFtE are to be calculatedespecially in very small babies with a rapid fR a samplingrate of 200 Hz is recommended

Acknowledgements The authors would like to thankall other members of the Task Force who contributed todeveloping these recommendations J Allan (Philadel-phia PA USA) E Bar-Yishay (Jerusalem Israel) CBeardsmore (Leicester UK) R Castile (Colombus OHUSA) JB Clough (Southampton UK) AL Coates(Toronto Canada) I Dundas (London UK) U Frey(Bern Switzerland) M Gappa (Hanover Germany) SGodfrey (Jerusalem Israel) I Goetz (London UK) RGregson (Southampton UK) P Gustafsson (SkovdeSweden) M Henschen (Freiburg Germany) A-F Hoo(London UK) A Jackson (Boston MA USA) J deJongste (Rotterdam the Netherlands) R Kraemer (BernSwitzerland) S Lum (London UK) P Merkus(Rotterdam the Netherlands) IT Merth (Leiden theNetherlands) M Morris (Little Rock AR USA) BReinmann (Bern Switzerland) P Seddon (BrightonUK) G Sharma (Chicago IL USA) M Silverman(Leicester UK) P Sly (West Perth Australia) RTepper (Indianapolis IN USA) D Vilozni (PetachTikva Israel) and E van der Wiel (Rotterdam theNetherlands) and all those around the world includingmembers of the industry who read the various draftsand provided valuable feedback

References

1 Frey U Stocks J Coates A Sly P Bates J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Specifications for equipment used forinfant pulmonary function testing Eur Respir J 2000 16731ndash740

2 Sly P Tepper R Henschen M Gappa M Stocks J onbehalf of the ERSATS Task Force on Standards for InfantRespiratory Function Testing Tidal forced expirationsEur Respir J 2000 16 741ndash748

3 Frey U Stocks J Sly P Bates J on behalf of the ERSATSTask Force on Standards for Infant Respiratory FunctionTesting Specifications for signal processing and data

33 36 39Time s

-75

0

75

0

300

-300

Flow

mLmiddots

-1

a)

b)

Volu

me

mL

Fig 8 ndash A a) 6-s segment near the end of the flow record shown infigure 7b and b) the corresponding volume signal The middle breathcontains several large oscillations in flow with peak magnitudes that areas large as (or larger than) those of true breaths in the recordConsequently the threshold algorithm cannot distinguish them from truebreaths The smoothed algorithm in contrast is successful at discardingthese oscillations because it smooths them out in a low-pass filteredversion of flow prior to looking for zero crossings to determineinspiratoryexpiratory transitions

1191TIDAL BREATH ANALYSIS IN INFANTS

handling used for infant pulmonary function testing EurRespir J 2000 16 1016ndash1022

4 Gappa M Colin AA Goetz I Stocks J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Passive respiratory mechanics theocclusion techniques Eur Respir J 2000 (in press)

5 Gappa M Colin AA Goetz I Stocks J on behalf of theERSATS Task Force on Standards for Infant RespiratoryFunction Testing Plethysmogrpahic measurements oflung volume and airway resistance Eur Respir J 2000 (inpress)

6 Morris MG Gustafsson P Tepper R Gappa M Stocks Jon behalf of the ERSATS Task Force on Standards forInfant Respiratory Function Testing The bias flownitrogen washout technique for measuring functionalresidual capacity Eur Respir J 2000 (in press)

7 Stick S Measurements during tidal breathing In StocksJ Sly PD Tepper RS Morgan WJ eds Infant RespiratoryFunction Testing 1st Edn New York John Wiley ampSons Inc 1996 pp 117ndash138

8 Schmalisch G Foitzik B Wauer RR Stocks J The in-fluence of apparatus deadspace on tidal breathing para-meters in newborns comparison of the flow-throughand conventional techniques Eur Respir J 2000 (inpress)

9 Perez W Tobin MJ Separation of factors responsible forchange in breathing pattern induced by instrumentationJAppl Physiol 1985 59 1515ndash1520

10 Fleming PJ Levin MR Goncalves A Changes in respi-ratory pattern resulting from the use of a face mask torecord respiration in newborn infants Pediatr Res 198216 1031ndash1034

11 Dolfin T Duffty P Wilkes D England S Bryan H Effectsof a face mask and pneumotachograph on breathing insleeping infants Am Rev Respir Dis 1983 123 977ndash979

12 Emralino F Steele AM Effects of technique and analyticconditions on tidal breathing flow volume loops in termneonates Pediatr Pulmonol 1997 24 86ndash92

13 Gaultier C Fletcher M Beardsmore C Motoyama EStocks J Measurement conditions In Stocks J Sly PDTepper RS Morgan WJ eds Infant Respiratory FunctionTesting 1st Edn New York John Wiley amp Sons Inc1996 pp 29ndash44

14 Brown K Aun C Jackson E Mackersie A Hatch DStocks J Validation of respiratory inductive plethysmo-graphy using the qualitativediagnostic calibrationmethodin anaesthetized infants Eur Respir J 1998 12 935ndash943

15 Bates JHT Turner MJ Lanteri CJ Jonson B Sly PDMeasurement of flow and volume In Stocks J Sly PDTepper RS Morgan WJ eds Infant Respiratory FunctionTesting 1st Edn New York John Wiley amp Sons Inc1996 pp 81ndash116

16 Morris MG A simple new technique to measure theeffective dead space of the face mask with a watervolumeter in infants Eur Respir J 1999 14 1163ndash1166

17 Farre R Montserrat JM Rotger M Ballester E NavajasD Accuracy of thermistors and thermocouples as flow-measuring devices for detecting hypopnoeas Eur RespirJ 1998 11 179ndash182

18 Renzi PE Giurdanella CA Jackson AC Improvedfrequency response of pneumotachometers by digitalcompensation J Appl Physiol 1990 68 382ndash386

19 Roske K Foitzik B Wauer RR Schmalisch G Accuracyof commercial equipment for ventilatorymeasurements inventilated newborns J Clin Monit 1998 14 413ndash420

20 van der Ent CK Brackel HJL Mulder P Bogaard JMImprovement of tidal breathing pattern analysis in child-ren with asthma by on-line automatic data processingEurRespir J 1996 9 1306ndash1313

21 Schmalisch G Foitzik B Wauer RR Stocks J In vitroassessmentof equipment and software used to assess tidalbreathing parameters in infants Eur Respir J 2000 (inpress)

22 Foitzik B Schmidt M Windstetter D Wauer RRSchmalisch G Leak measurements in spontaneouslybreathing premature newborns by using the flow-throughtechnique J Appl Physiol 1998 85 1187ndash1193

23 Dundas I Dezateux CA Fletcher ME Jackson EAStocks J Comparison of single-breath and plethysmo-graphic measurements of resistance in infancy Am JRespir Crit Care Med 1995 151 1451ndash1458

24 Schmidt M Foitzik B Wauer RR Winkler F SchmalischG Comparative investigation of algorithms for the de-tection of breaths in newborns with disturbed respiratorysignals Comp Biomed Res 1998 31 413ndash425

1192 JHT BATES ET AL