8/8/2019 Pub 012

1/7

Reprinted with permission fromPACINGND CLINICALLECTROPHYSIOLOGY,olume 17, NO. 8, August 1994Copyright O 1994 by Futura Publishing Company, Inc., Armonk, NY 10504-0418

Digital Signal Processing Chip Implementationfor Detection and Analysis of IntracardiacElectrogramsCHIH-MING JAMES CHIANG, JANICE M. F NK IN S,and LORENZO A. DICARLOFrom the Department of Electrical Engineering and Computer Science, College of Engineering,University of Michigan, and The Michigan Heart and Vascular Institute and CardiacElectrophysiology Laboratory, St. Joseph Mercy Hospital, Ann Arbor, Michigan

CHIANG, C.-M.J., ETAL.: Digital Signal Processing Chip Implementation for Detection and Analysis ofIntracardiac Electrograms. The adoption of digital signal processing (DSP) microchips for detection andanalysis of electrocardiographic signals offers a means for increased computational speed and the oppor-tunity for design of customized architecture to address real-time requirements. A system using the Moto-rola 56001 DSP chip has been designed to realize cycle-by-cycle detection (triggering) an d waveformanalysis using a time-domain template matching technique, correlation waveform analysis (CWA). Thesystem digitally samples an electrocardiographic signal a t 1000 Hz, incorporates an adaptive trigger fordetection of cardiac events, an d classifies each waveform as normal or abnormal. Ten paired sets ofsingle-chamber bipolar intracardiac electrograms (1-500 Hz) were processed with each pair containinga sinus rhythm (SR) passage a nd a corresponding arrhythmia segment from the same patient. Four of tenpaired sets contained in traatrial electrograms that exhibited retrograde atrial conduction during ventricu-la r pacing; the remaining six paired sets of intraventricular electrograms consisted of either ventriculartachycardia (4) or paced ventricular rhythm (2). Of 2,978 depolarizations in the test set, the adaptivetrigger failed to detect 6 (99.8% detection sensitivity) an d had 11false triggers (90.6% specificity). Usingpatient dependent thresholds for CWA to classify waveforms, the program correctly identified 1,175 of1 I97 (98.2% specificity) sin us rhythm depolarizations and 1,771 of 1,781 (99.4% sensitivity) abnormaldepolarizations. From the results, the algorithm appears to hold potential for applications such a s real-time monitoring of electrophysiology studies or detection and classification of tachycardias in implantableantitachycardia devices. (PACE 1994; 17:1373-1379)arrhythmia, intracardiac electrogram, antitachycardia device, implantable defibrillator

Introduction amplitude, area under the curve) were advancedfor diagnostic electrocardiograms (ECGs).' AsMor~hological Of cardiac waveforms computer techniques became morehas long been a in better methods were used, particularly in the clas-d i O g r a ~ h ~ . measures sification of and abnormal depolarizationsfor such applications as mal-time coronary caremonitorine and fast-time scanning of ambulatory"This work was partially supported by National Science Foun- (Halter) monitoring. The method choice in thesedation Grants BCS-8909042 and EID-9023514. applications converged in the late 1970s to a point-Address for reprints: Chih-ming J Chiang. Ph.D.. Telectmnics bypoint comparison of each digitized waveformPacing Systems, 7400 S. Tucson Way, Englewood, CO 80112.Fax: (303) 799-2213. with a previously extracted template of a normal

Received June 25, 1993; revision January 10, 1994; accepted d e p ~ l ~ i z a t i o n . ~ ~ ~his technique used the classicFebruary 2, 1994. statistical correlation coefficient? and later be-

PACE, Vol. 1 7 August 1994 1 3 7 3

8/8/2019 Pub 012

2/7

CHIANG, ET AL.

came known as the correlation waveform analysis Adjustable Threshold Trigger(CWA].5-7CWA is a robust measure; it is imper-vious to the DC offset of the ECG and to amplitudevariations of the waveforms. More recently, CWAhas been applied to morphological analysis of in-tracardiac electrograms for recognition of abnor-mal activation recorded by catheter electrode^.^ Ithas been shown to be equally robust in this setting,given that appropriate bandwidths of the recordedsignal be ob~erved.~ue to the heavy computa-tional demands of CWA, however, its only real-time application has been limited to coronary caremonitoring systems.One application that would benefit from in-corporation of morphological methods are anti-tachycardia devices such as pacing and nonpacingimplantable cardioverter defibrillators. These de-vices are designed to terminate potentially life-threatening tachyarrhythmias such as ventriculartachycardia (VT) and ventricular fibrillation [VF)and have been shown to perform well in prevent-ing sudden cardiac death.' However, these-devicesutilize simple detection schemes based upon rateand yield a high incidence of false delivery of ther-apy due to the one-channel nature of detection,which frequently confuses supraventricular ar-rhythmias with VT and VF. It has been suggestedthat morphological analysis is the key to alleviatethis p r ~ b l e m . ~ ? ~o date, however, the only real-time programs using CWA for intracardiac elec-trogram waveform a n a l y s i ~ ~ ~ ~ ~re implementedon the PC-based 386 computer (Intel P700, IntelCorp., Beaverton, OR, USA) or coronary care moni-toring machines.

The goal of this study was to demonstrate thefeasability of implementing CWA on a digital sig-nal processing (DSP) chip (Motorola 56001 DSP,Motorola Communications & Electronics, Inc.,Schaumburg, IL, USA). This would represent afirst step towards specialized architecture incor-porating CWA that would be utilized in future gen-eration implantable antitachycardia devices.

Methods

An accurate triggering mechanism is essen-tial for rate and morphological analysis of intra-cardiac electrograms. A previously developedmultistage software depolarization detector usingadjustable thresholds has been shown to be ro-bust in detecting intracardiac electrogramsl' andis the basis for the trigger implemented on theDSP chip.' The scheme contains a second order IIR But-terworth band-pass filter that suppresses unde-sired frequency components and eliminates base-line shift. The trigger has an adaptable thresholdwith exponential time decay to allow for variabil-ity in waveform amplitude. A blanking period isimposed to prevent multiple event detection on asingle depolarization.

Figure 1 shows a comprehensive diagram ofthe trigger. The first stage of the filter is a digitalband-pass IIR filter H(z) derived from the bilinear

Bandpass2060Hz

Triggera dBlank tornext 150mSThe DSP program contains main Figure 1. The block diagram for the trigger. x = input,subsections, the first a depolarization detection = outputfrom he bandpassfilter,w = the magnitudescheme described in the "Adjustable Threshold ,fy, = fraction to multiply with W, d = fraction toTrigger" subsection, and the second the CWA de- multiply with previous vvalue, v = value from Equationscribed in the CWA subsection. 2,Max = maximum.

1374 August 1994 PACE, ol. 17

8/8/2019 Pub 012

3/7

DSP CHIP IMPLEMEN TATION OF CORRELATION WAVEFORM ANALYSIS

transform of a second order analog filter H(s) with Hz, fH = 60 Hz, d = 2-1'1000, and r = 0.5 baseds = (1 - - l)I(l + zF1)ll: upon empirical results.ll

and ao, a,, a, are derived as follow^'^:

where WL, W H re:

with f~ he lower cut-off frequency, fH the highercut-off frequency, and f, the sampling frequencyof the system. The resulting difference equation is:Yi = 00 (xi - xi-,) - alyj-1 - a~yi-, (I)

where yi and xi are the ith filter output and input,respectively.The second stage is a threshold comparison ofthe magnitude of the filtered signal with an adjust-able threshold. The threshold v is the maximumof two components, the first a fraction (r) of therectified signal and the second a decay factor (dwith d < 1) multiplied with the previous thresh-old. The update function for the threshold is:

The two factors, rand d, control the sensitivity ofthe trigger to allow for detection of waveformswith varying amplitudes. Making r small and dsmall will increase sensitivity of detection at theexpense of increased false triggering probability.The third stage is a blanking interval that preventsmultiple triggering during a single depolarization.For this study, values were set as follows:fL = 20

CWAThe CWA performmce measure p, indepen-dent of amplitude fluatuations and baselinechanges, yields an output between - 1 and + 1where + 1 indicates a perfect match. Mathemati-cally, p is defined as1':

where N = the number of points in the template;ti = the template points; si= the signal points tobe processed; t = the template average; and s =the signal average.To further reduce computation, we eliminatethe square root calculation, and rename the met-ric C3 = sign(p)$ where sign is f 1 dependingon the sign of p. If normalization of the templatepoints has been calculated (such as by an externalprogrammer) in advance such that i = 0 andy t i = 1, hen the on-line calculation of newlyintroduced depolarizationg becomes12:

[Z jI (ti) si))2(4)

This represents a streamlined version of CWA(using anN point template), which requires 2N +2 multiplications, 1 division, and 3N + 1 addi-tions per waveform. This is the equation that isimplemented on the DSP chip.Motorola 56001 System

PACE, Vol. 17 August 1994

The Motorola 56001 DSP chip was designed toperform fast multiply and accumulate operationsappropriate for DSP procedures, such as filteringand Fast Fourier Transforms (FFT).13.14The chipis based on a modified Harvwd architecture13 witha 512 word dynamic random access program mem-ory (DRAM) and two on-chip X and Y data read-only-memory (ROM) as well as data random accessmemory (RAM) (256 in X and 256 in Y for bothRAM and ROM).13.14 t has a 24-bit data register

8/8/2019 Pub 012

4/7

CHIANG, ET AL.

and two accumulators of 56-bit precision. It exe-cutes in fixed point arithmetic with a computa-tional speed of 10.25 million instructions per sec-ond (MIPS).13.14 he system contains a 16-bit ana-log to digital (AID) converter with a maximumsampling rate of 44 ~ H z . ' ~ , ' ~The implemented DSP program contains threemain components, the analog to digital interface,the triggering section, and the CWA section. Essen-tially, the algorithm remains in a wait state untilan interrupt occurs indicating that data have beenreceived from the AID. Upon this event, the pro-gram stores the data and enters the triggering sec-tion. If a depolarization occurs, CWA is activated.CWA is implemented using Equation 4 to reducecomputational load. Template values are normal-ized prior to utilization and a 64 point CWA (64msec at a 1,000Hz sampling rate) was chosen sincemost intracardiac waveforms fall within this dura-tion and since 64 is a power of 2 (z6).To align thetemplate with the waveform being analyzed, thewindow centered around the trigger for CWA cal-culation was shifted by n points, with n varyingfrom - 10 to +10, creating 21 separate windows.CWA was performed for each of the windows andthe maximum value of the 2 1 taken as the trueCWA performance measure. If this measure wasgreater than a patient dependent threshold of 0.9or 0.81 (since signal amplitudes and variabilitieswere patient dependent, the threshold were cho-sen individually for each pair of passages), the DSPchip output a positive pulse for a normal depolari-zation, otherwise it gave a negative pulse indicat-ing an abnormal waveform.

MaterialsTen pairs of single channel bipolar intracar-diac electrograms (1-500 Hz) from seven patientswere processed with each pair consisting of onesinus rhythm (normal) passage and one corre-sponding arrhythmia segment. Four of the pairswere intraatrial electrograms containing retro-grade conduction during ventricular pacing. The

other six pairs were intraventricular electrogramscontaining ventricular tachycardia (4) or ventricu-lar pacing (2).Data used for analysis were recorded from 6-French quadrapolar electrode catheters (USCI Di-vision, C.R. Bard Inc., Billerica, MA, USA) withan interelectrode distance of 1cm. The recordings

were made in the high right atrium and right ven-tricular apex on FM magnetic tape (Hewlett-Pack-ard Model 3968A, San Dieglo, CA, USA) at 3:inches per second during prov~cat ive lectrophys-iology studies with patients lying supine. For pro-cessing, data were digitized through the DSP AIDat 1,000 Hz and subjected to analysis on the DSPchip.

ResultsOf the 2,978 total depolaL.izations from the 10pairs studied, the adaptive trigger missed only 6(99.8% sensitivity of depolmization detection)while giving11false-positives (99.6% specificity).Results are shown in Table I. For the performanceof CWA with patient dependent thresholds and

shifting window calculations (seeMotorola56001System), the program correctly identified 1,171 of1,197 (97.8% specificity) sinus rhythm (normal)depolarizations and 1,771 of 1,781 (99.4% sensi-tivity) abnormal depolarizations. The overall re-sult was 2,943 of 2,978 (98.8%) (Table 11).Figure 2 contains intraventricular electrogramresults from patient 4. The lleft side of the figuredepicts a sinus rhythm pass ge along with the di-agnostic signals and the right side has a VT seg-ment with its correspondin$ program diagnosis.As can be seen from the sinus rhythm passage, trig-ger and CWA sections functioned properly withthe periodic positive pulse outputs indicating nor-mal waveform diagnosis. For the VT passage, thetrigger works equally well and the negative pulsereflects abnormal waveform diagnoses.Figure 3 contains intraatrial electrogram re-sults from a sinus rhythm passage followed by aventricular pacing section (yielding retrogradeatrial activation) with its corresponding diagnosticpulses. There are no false trliggers and all normaldepolarizations are indicated in the event markerby positive pulses and abnormal ones by negativepulses. The black blotches represent discontinu-ities during the taping process and their corre-sponding results are disregwded when tabulatingsystem performance.Figure 4 contains an intraatrial electrogrampassage that revealed triggering failures. The be-ginning of the trace shows dnus rhythm followedby ventricular pacing. Note the false-positive trig-gering in the pacing segment due to a wandering

1376 August 1994 PACE, Vol. 17

8/8/2019 Pub 012

5/7

DSP CHIP IMPLEMENTATION OF CORRELATION WAVEFORM A NALYSIS

Table I.Result for Real-Time lm~lementation f the Triaaer

Sinus Rhythm ArrhythmiaPatient Channel Num FP FN Num FP FN

123456789

10Total

Grand Total- - - ---um = number; FP = false-positive; FN = false-negative.baseline. CW A also failed in one depolarization intracardiac electrograms, The trigger obtained anduring sinus rhythm when a normal waveform accuracy of 99.4%.The few cases of missed trig-was declared abnormal. gering (612,968) esulted from a sequence in whicha large amplitude waveform was followed by one

Discussion andConclusion of small amplitude, such that the threshold re-mained too high for the second depolarization toAs shown previously, the DSP system per- be detected. ~aise-positiveriggers occurred due toformed well in waveform detection and CWA on wandering baseline coupled with small amplitude-

Table 11.Result for Real-Time lm~lementation f CWA

Normal AbnormalPatient Channel Num Wrong Num Wrong

1 V 170 13 105 02 V 84 2 89 03 V 191 1 271 64 V 137 0 214 05 A 146 7 249 06 A 70 0 229 07 A 59 0 143 08 V 50 1 134 09 A 164 0 156 4

10 v 126 - 191 -Total 1,197 26 1.781 -0Grand Total 2,978 36CWA = correlation waveform analysis; Num = number.

PACE, Vol. 17 August 1994

8/8/2019 Pub 012

6/7

CHIANG, T AL.

Ventricular Electrogram

SINUS RHYTHM VENTRICULAR TACHYCARDIA

DSP OutputFigure 2 . Patient 4 intraventricular electrograms are shown. The left side of the figure containsa sinus rhythm passage with accompanying diagnostic signals and the right side has a ventriculartachycardia passage along with the diagnosis.

signals. The small incidence of these events doesnot warrant modification of the triggering mecha-nism. A previous study examining the perfor-mance of this depolarization detection schemeshowed that the trigger performed well." TheMedtronic PCD (Medtronic, Inc., Minneapolis,MN, SA) detection scheme is also similar to thetrigger presented in this study.For CWA measures, the program was able to

distinguish between normal and abnormal wave-forms with at least 98% accuracy using patient de-pendent thresholds. Most errors were normalwaveforms classified as abnormal due to low CWAvalues and these were mainly due to inaccuracyof the trigger location. The 21-window CWA calcu-lation alleviates most of the problem of inexact-ness of depolarization detection.In conclusion, a real-time program has been

Atrial Electrogram

SINUS RHYTHM VENTRICULAR PACING

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 l l l l l l l l l l l l ~ l l l l

DSP OutputFigure 3. Intmatrial electrogram results. Sinus rhythm is followed by pacing.

August 1994 PACE, Vol. 17

8/8/2019 Pub 012

7/7

DSP CHIP IMPLEMENTATION OF CORRELATION WAVEFORM ANALYSIS

Atrial Electrogram

VENTRICULAR PACING

/ ' I l l 1 I L L l l l l l 1 1 1 1 I I ~ I I I I I I I . I ~ ~ ~ I I I IA A AAA AADSP Output False Positives (FP)

Figure 4. An intraatrial passage with sinus rhythm followed by ventricular pacing. This showshow the wandering baseline corrupts the trigger peqormance with false-positive (FP) detectionof waveforms. Notice also the incorrect digital signal processing (DSP) output of negative pulseon sinus rhythm passage due to low correlation value.

implemented on the Motorola56001DSP chip that miniaturization that would make incorporation ofdetects cardiac activation and performs CWA on CWA possible in real-time applications such asintracardiac electrograms. The program performs next-generation implantable antitachycardia de-well and represents an important first step towards vices.References

Jenkins J. Automated electrocardiography and ar-rhythmia monitoring. Prog Cardiovasc Dis 1983;25:367-408.Arzbaecher R, Biancalana P, Stibolt T, et al. Com-puter technique for detection of cardiac arrhyth-mias. (abstract) J Assoc Advance Med Instrument1971; 5:104.Feldman C, Arnazeen P, Klein M, et al. Computerdetection of ventricular ectopic beats. Comput Bio-med Res 1971; 4:666-674.Kennedy JB, Neville AM. Basic statistical methodsfor engineers and scientists. Harpers & Row, Pub-lishers, New York, NY, 1986, pp. 410-411.Lin D, DiCarlo L, Jenkins J. Identification of ven-tricular tachycardia using intracavitary ventricularelectrograms: Analysis of time and frequency do-main patterns. PACE 1988; 11:1592-1605.DiCarlo L, Throne R, Jenkins J. A time-domainanalysis of intracardiac electrograms for arrhyth-mia detection. PACE 1991; 14:329-336.Throne R, Jenkins J, Winston S, et al. Discrimina-tion of retrograde from anterograde atrial activa-tion using intracardiac electrogram waveformanalysis. PACE 1989; 12:1622-1630.Cohen T, ChienW , Lurie K, et al. Implantable car-

dioverter defibrillator proarrhythmia: Case reportand review of the literature. PACE 1991; 14:1326-1329.9. Chiang CJ, Jenkins JM, DiCarlo LA. An innovativetwo-channel rate and morphology method forcomplex arrhythmia diagnosis in implantable de-vices. 14th Annual International Conference of theIEEE Engineering in Medicine and Biology Soci-ety, 1992;496-497.10. Chiang CJ, Jenkins JM, DiCarlo LA. Real-time ar-rhythmia identification from automated analysisof intraatrial and intraventricular electrograms.PACE 1993; 16(1):223-227.11. McDonald R, Jenkins J, Arzbaecher R, et al. A soft-ware trigger for intracardiac waveform detectionwith automatic threshold adjustment. IEEE Com-puters in Cardiology, IEEE Press, New York, NY,1989, pp. 167-170.12. Throne RD.Analysis of template matching meth-ods for intracardiac electrogram analysis. Ph.D.Thesis, University of Michigan, 1990.13. Lee E. Programmable DSP Architectures: Part I.IEEE ASSP Magazine 1988; 5(4):4-19.14. Lee E. Programmable DSP Architectures: Part 11.IEEE ASSP Magazine 1989; 6(1):4-14.

PACE, Vol. 17 August 1994 1379