D.V. Moretti et al., 2003

International Journal of Psychophysiology 47(2003) 199–216

0167-8760/03/$ - see front matter� 2002 Elsevier Science B.V. All rights reserved.PII: S0167-8760Ž02.00153-8

Computerized processing of EEG–EOG–EMG artifacts for multi-centric studies in EEG oscillations and event-related potentials

D.V. Moretti *, F. Babiloni , F. Carducci , F. Cincotti , E. Remondini , P.M. Rossini ,a, a a,b a a b,c,d

S. Salinari , C. Babilonie a,b

Dipartimento di Fisiologia Umana e Farmacologia, Sezione di EEG ad Alta Risoluzione,a

Universita degli studi di Roma ‘‘La Sapienza’’, P.le Aldo Moro, 5, 00185 Rome, Italy`IRCCS ‘‘S. Giovanni di Dio-F.B.F.’’, Brescia, Italyb

A.Fa.R. IRCCS Div. di Neurologia, Osp. FBF Isola Tiberina, Rome, Italyc

Clinica Neurologica, Universita ‘‘Campus Biomedico’’, Rome, Italyd `Dip. di Informatica e Sistemistica, Universita di Roma La Sapienza, Rome, Italye `

Received 4 June 2002; received in revised form 22 October 2002; accepted 29 October 2002

Abstract

The aim of this study is to present a package including standard software for the electroencephalographic(EEG),electro-oculographic(EOG) and electromyographic(EMG) preliminary data analysis, which may be suitable tostandardize the results of many EEG research centers studies(i.e. multi-centric studies) especially focused on event-related potentials. In particular, our software package includes(semi)automatic procedures for(i) EOG artifactdetection and correction,(ii) EMG analysis,(iii ) EEG artifact analysis,(iv) optimization of the ratio between artifact-free EEG channels and trials to be rejected. The performances of the software package on EOG–EEG–EMG datarelated to cognitive–motor tasks were evaluated with respect to the preliminary data analysis performed by two expertelectroencephalographists(gold standard). Due to its extreme importance for multi-centric EEG studies, we comparedthe performances of two representative ‘regression’ methods for the EOG correction in time and frequency domains.The aim was the selection of the most suitable method in the perspective of a multi-centric EEG study. The resultsshowed an acceptable agreement of approximately 95% between the human and software behaviors, for the detectionof vertical and horizontal EOG artifacts, the measurement of hand EMG responses for a cognitive–motor paradigm,the detection of involuntary mirror movements, and the detection of EEG artifacts. Furthermore, our results indicateda particular reliability of a ‘regression’ EOG correction method operating in time domain(i.e. ordinary least squares).These results suggest that such a software package could be used for multi-centric EEG studies.� 2002 Elsevier Science B.V. All rights reserved.

Keywords: Preliminary EEG data analysis; EOG artifacts correction

*Corresponding author. Tel.:q39-6-49910989; fax:q39-6-49910917.E-mail address: [email protected](D.V. Moretti).

200 D.V. Moretti et al. / International Journal of Psychophysiology 47 (2003) 199–216

1. Introduction

Conventional electroencephalography(EEG) iscommonly used in the evaluation of some neuro-logical diseases, such as epileptic syndrome, met-abolic encephalopathy and cerebral parenchymainfective diseases(Daly and Pedley, 1990; Hughes,1994). Recently, several computational techniqueshave been proposed to estimate EEG corticalsources and to quantify the event-related changesof EEG oscillations(Urbano et al., 1996; Babiloniet al., 1998, 1999; Neubauer et al., 1999; Pfurt-scheller and Lopez da Silva, 1999; Babiloni et al.,2000). These techniques are promising for clinicalapplications(Leuchter et al., 1993; Locatelli et al.,1998; Aurlien et al., 1999).A correct preliminary EEG–EOG–EMG data

analysis is especially important in psychophysio-logical settings in which attentional and cognitivedemands can increase involuntary movements aswell as the number and types of EOG artifactsover frontal scalp regions. It is then clearly impor-tant that the detection andyor correction of theseartifacts(especially EOG) are effective and stan-dardized in a multi-centric study, to avoid errone-ous inferences on the role of prefrontal cortex.Anyway, some methodological difficulties pose

obstacles to the development of multi-centric stud-ies, aimed at creating wide EEG databases to beused in clinical practice. One of the major obsta-cles is a great inter-rater variability in the prelim-inary data analysis, i.e. the detection ofinstrumental or biological(i.e. EOG or EMG)artifacts affecting EEG signals. The utilization ofqualitative criteria in the preliminary data analysisincreases the variance of final results and makesimpossible or limits the standardization of distinctresearch unit data for meta-analysis studies. Ideally,the inter-rater variability in the EEG preliminarydata analysis should be removed by a unifiedsoftware package that is easy to use.For a multi-centric EEG study, an important

phase of the preliminary data analysis is thedetection and correction of the EOG artifacts,especially when EEG experiments are carried outin neurological patients or in subjects involved insensorimotor or cognitive demands lasting severalseconds. In literature, in order to use the majority

of recorded EEG single trials, two main approachesfor the detection and correction of EOG artifactshave been proposed. The first approach estimatesthe coefficients modeling the transmission of sig-nals from EOG to single EEG channels by ‘regres-sion’ methods(Croft and Barry, 2000a). Thesemethods provide a sufficiently reliable EOG cor-rection in both time(Hillyard and Galambos, 1970;Verleger et al., 1982; Woestenburg et al., 1983)and frequency(Whitton et al., 1978; Woestenburget al., 1983) domains. The second approach isbased on the principal component analysis(PCA;Lagerlund et al., 1997), independent componentanalysis(ICA; Bell and Sejnowski, 1995; Makeiget al., 1999), and dipole modeling(Berg andScherg, 1991; Lins et al., 1993a,b; Berg andScherg, 1994). These ‘spatial component’ methodsderive spatial filters or maps that decompose theEEG topographical distribution into a sum ofspatial components. With this approach, the EEGspatial components due to eye movements orblinks are recognized as distinct by those due tobrain sources and are removed from the recordedEEG distribution. In particular, PCA and ICA mayrequire manual recognition of the spatial compo-nents of EEG data supposed to be due to eyemovements. In contrast, the dipole approach forEOG artifact correction is a fully automatic pro-cedure, based on an explicit mathematical model-ing of the eyes as EOG artifact sources and of thehead as a volume conductor(Berg and Scherg,1994). On the whole, the ‘regression’ methodsrepresent a well-known approach easy to be imple-mented and used(Croft and Barry, 2000b), while‘spatial component’ methods are ideal tools forbasic research purposes being very efficient in theEOG correction(Picton et al., 2000a).The aim of this study is to present the perform-

ances of a package including standard software forthe EEG, EOG and EMG preliminary data analy-sis, which may be suitable to standardize theresults of multi-centric EEG studies. This is adesirable target, but there is poor literature evi-dence in favor of that. The general idea at thebasis of the present study is that this target couldbe reached if EEG community provides much moreevidence on the performances(i.e. pros and cons)of the computer programs for the preliminary

201D.V. Moretti et al. / International Journal of Psychophysiology 47 (2003) 199–216

EOG–EEG–EMG data analysis. The present studyrepresents an experimental contribution towardsthat target.In particular, our software package includes

(semi)automatic procedures for(i) EOG artifactdetection and correction,(ii) EMG analysis,(iii )EEG artifact analysis,(iv) optimization of theratio between artifact-free EEG channels and trialsto be rejected. The performances of the softwarepackage were evaluated with respect to the prelim-inary data analysis performed by two expert elec-troencephalographists(gold standard). Due to itsextreme importance for multi-centric EEG studies,we compared the performances of two representa-tive ‘regression’ methods for the EOG correctionin time and frequency domains. The aim was theselection of the most suitable method in the per-spective of a multi-centric EEG study.

2. Materials and methods

2.1. Overview of the software package

The software was developed inMATLAB 5.1environment, based mainly on criteria and algo-rithms already accepted by the Scientific Com-munity. However, it runs properly even inMATLAB6.5.The main functions of the software are the

following; (i) statistical recognition and removalfrom the original directory of the EEG single trialsassociated with involuntary mirror EMG activityduring the motor task(unilateral voluntary motortask); (ii) recognition (50-mV threshold) of theEEG single trials contaminated by ocular move-ment or blinking and subsequent removal from theoriginal directory; (iii ) correction of EEG singletrials based on EOG artifact morphology(blinksor saccades); (iv) listing of the recordings channelsand EEG single trials free from artifacts. Of note,the operator can indicate the minimum number ofEEG single trials necessary. An optimization pro-cedure discards the most artifactual recordingchannels up to fit the minimum number of singletrials required for further analysis;(v) estimationof the onset, offset and duration of EMG response(operant hand) related to voluntary movements.

2.2. Detection of EMG activity not-task-related(non-operant hand)

The detection is carried out on the EMG dataof the non-operant hand(i.e. the hand not involvedin the motor task).For each single trial, the software splits EMG

activity (non-operant hand), in segments lasting500-ms each. In two contiguous EEG segments,EMG activity can be reasonably supposed to bestationary. Subsequently, Student’s pairedt-test isapplied to compare mean amplitude of contiguousEMG activity segments. The null hypothesis isthat the mean amplitude of contiguous EMG activ-ity segments does not differ statistically(P)0.05).The EMG single trials where the null hypothesisis rejected are considered as artifactual(i.e. withmirror movements; Fig. 1).

2.3. Detection of the onset, offset, duration, andarea of EMG task-related response (operant hand)

The detection is carried out on the EMG dataof the operant hand(i.e. the hand involved in themotor task).The software operates as follows(Fig. 2): (i)

estimation of a first threshold(‘noise’), defined asthe mean EMG activity(amplitude) value plusthree standard deviations, in the first second of therecording(absolute baseline); (ii) measurement ofthe EMG peak latency and the amplitude relatedto the movement(zerotime). EMG peak is definedas the instant reaching the maximum amplitude ofthe EMG potential;(iii ) recognition of a timeperiod(period of interest) that ranges fromy500to q500 ms with respect to the zerotime;(iv)within the period of interest, the software computesother two amplitude thresholds, defined as 45 and90% of the EMG peak amplitude(baseline abso-lute value as a reference). Moreover, in the periodof interest, the software defines four contiguousperiods of 15 ms. During the first of them, areference baseline(relative baseline) is estimated.The EMG onset time is defined as the instant afterwhich amplitude values in the three subsequentperiods exceed three standard deviations of therelative baseline period. Furthermore, the onsetinstant must exceed the amplitude of all mentioned


Fig. 1. EMG activity in the operant and non-operant hand in arepresentative subject. The subject performed in a voluntarymotor task(right finger extension). The software divides EMGactivity (non-operant hand), of each trial into 500-ms seg-ments. In these segments, EMG activity can be reasonably sup-posed to be stationary. Subsequently, Student’st-test is appliedto compare mean amplitude of contiguous EMG activity seg-ments. The null hypothesis is that the mean amplitude of con-tiguous EMG activity segments does not differ statistically(P-0.05). The trials where the null hypothesis is rejected areconsidered as artifactual.

Fig. 2. Detection of EMG onset and offset. The proposed soft-ware recognizes the onset and offset of EMG, based on twosteps. In the first step, the software sets three amplitude thresh-olds related to the background noise(absolute baseline) as wellas to the partial and complete synchronization of the motorunits involved in the motor performances(45 and 90% ofEMG peak amplitude, respectively). In the second step,(seethe inset), the software computes four short(15 ms) EMGperiods in the temporal sequence. The first of period serves asa reference period(relative baseline) for the following threeperiods. The EMG onset time is defined as the instant afterwhich amplitude values in the three subsequent periods exceedthree standard deviations of the relative baseline period. In thesame time the signal has to exceed all the amplitude thresholds(‘noise’, 45, 90% of the EMG peak). Symbols of the figure:S1: background noise threshold; S2: 45% EMG peak ampli-tude; S3: 90% EMG peak amplitude; T1: time lasting betweenS1 and S2; T2: time lasting between S2 and S3; A1: amplitudedifference between S2 and S1; A2: amplitude differencebetween S3 and S2; V1: ‘lift’ rate of the EMG signal in thefirst tract (S1–S2); V2: ‘lift’ rate of the EMG signal in thesecond tract(S2–S3). The ‘lift’ rate is computed as follows,using a linear approximation: V1sA1yT1 and V2sA2yT2.The drop rate is computed in the same way.

thresholds(‘noise’, 45 and 90% of the EMG peak).If these conditions are not satisfied, the softwaremakes a shift of 15 ms towards the zerotime andit processes other four periods. The same procedureis used to determine the offset of EMG response.The software starts fromq500 ms and comesback across the EMG time series.

2.4. Detection, classification and removal of EOGartifacts

The software uses a standard threshold of 50mV (Verleger, 1993) for the detection of artifactsin horizontal and vertical EOG, to fit the standard-ization need of a multi-centric EEG study. The

trials exceeding the threshold(i) only in thevertical EOG,(ii) only in the horizontal EOG, and(iii ) in both vertical and horizontal EOG are


automatically disposed in different folders for thesubsequent EOG correction procedure. Of course,the experimenter can manually monitor the sortingprocedure, if the extension of the data set allowsit.

2.5. Methods for EOG artifacts correction

2.5.1. Ordinary least squareOrdinary least square(OLS) is a representative

‘regression’ method for the correction of EOGartifacts in time domain(Verleger et al., 1982;Gratton et al., 1983; Jerwis et al., 1985; Croft andBarry, 1998a,b, 2000a,b,c). OLS method estimatesin an autoregressive manner the parameters of themathematical model that describe the recordedEEG as an overlap of the real EEG and vertical(andyor horizontal) EOG channel. These parame-ters are estimated minimizing the mean squareerror of the autoregressive model and can beapplied trial-by-trial, electrode-by-electrode oreven, as in the present variant OLS, segment-by-segment of a trial. This iterative procedure pro-vides an estimation of the transmission coefficientsfor all EEG trials.In particular, OLS is formally described by

y(i)su x qu x qu x qu x qe(i)1 1 2 2 3 3 p pn

s u x (i)qe(i)j j8js1

is1,2,«,m

where i is a generic sample of recorded EEGchannel recordedy(i), x (i) is the portion of EEGj

data due to EOG artifacts,e(i) is the EEG activity,m is the number of the samples for each trial andn is the order of the model, not necessarily equalto m.The equation can be written as follows:

YsXQqE;

whereY, X, Q andE are, respectively

w xYs y(1)y(2)y(i)∆y(m) ;T T T Tw xXs x (1)x (2)x (i)∆x (m) ;

w xQs u u ∆u ;1 2 p

w xEs e(1)e(2)e(i)∆e(m) ;

whereY is a column vector withm components,X is a dimensionwm, px matrix with x is a rowT

vector with p components,Q is a column vectorwith p components andE is a column vector withm components. For the estimation of the modelparameters, OLS utilizes the following equation:

y1T TQs X X X Y.Ž .

Noteworthy, OLS presumes that the samples ofEEG signals are not related. This means that:

2w xE e(i),e( j) ss isjs0 i(j

wheres is the EEG signal variance.2

In this condition, OLS estimation may beimproved with an autoregressive EEG model asfollows:

n

e(i)s f e(iyj)qa(i)j8js1

wheref are the model parameters anda(i) thej

non-related sequence.We used a third order model that results in:

3

y9(i)sy(i)y f y(iyj)j8js1

3T Tx9(i)sx (i)y f x (iyj)j8

js1

which gives:

Ty9(i)sx9 (i)Qqa(i).

The estimation of EOG artifacts correction par-ameters by OLS can be summarized as follows:(i) OLS provides a first estimation of theQ vectorand ofF vector; (ii) based on the previous twoestimates, OLS computesy9(i) and x9(i), whichallow a more refinedQ estimation and a newestimation of the vectors;(iii ) OLS repeats theestimation up toF vector estimation is convergent;


(iv) OLS uses the last estimation ofQ for theEEG correction.

2.5.2. Frequency domain correctionFrequency domain correction(FREQ) has been

previously described in detail(Van den Berg-Lenssen et al., 1994). In the set of single trials,we selected five 256-sampled intervals on EEGand EOG channels, based on the minimum andthe maximum EOG amplitude. A 256-sampledzero padding was carried out in all trials. FFT wasthen applied to every epoch selected. If we denotethe FFT of the maximum power EOG parts withpedex p and that with the minimum EOG powerwith pedex np (EEG(v) EOG(v) EEG(v)p p np

EOG(v) ), the transmittance coefficient of EOGnp

artifact for each EEG channel is computed as:

b(v)sU UEEG(v) EOG (v) y EEG(v) EOG (v)Ž . Ž .p p np np8 8

UEOG(v) EOG (v)Ž .p p8

where the range for the summations(from 1 tothe number of epochs selected for each channel)and * show the conjugated complex. In this for-mula, the denominator contains the mean perio-dogram of the EOG on the five 256-sampledintervals (maximum power, normalization factorT to get the power). The numerator contains the2

subtraction of the EEG–EOG mean cross-perio-dograms for the EOG maximum and minimumpower epochs. Periodograms are smoothed beforethe division and the temporal sequences are mul-tiplied by a cosine function before FFT application.Of note, an optimal performance of the methodwould have required the use of a transfer functionaveraged by several subjects for each frequencyband. However, this was beyond the scope of thepresent study, which was focused on the perform-ance of a basic FREQ procedure quick and easyfor the purposes of a large multi-centric study.

2.6. EEG artifact detection

2.6.1. Statistical methodEach EEG single trial is divided in 500-ms

lasting periods. Student’s pairedt-test is thenapplied to compare mean amplitude of contiguous

EEG periods. The null hypothesis is that the meanamplitude of the EEG signal is stationary incontiguous periods of 500 ms. Artifactual EEGchannels are those in which the test finds statisticalsignificant(P-0.05) differences.

2.6.2. Threshold methodThis standard method considers as artifactual

the EEG channels where the signal amplitude isover a fixed threshold(i.e. 100mV).

2.7. Optimization ratio between the number ofrecording channels and free-artifact trials

The software(Fig. 3) lists the numbers of theartifact free EEG single trials. If the total numberof single trials is not high enough for the furtherdata analysis, the software requires as an input theindication of the minimum number of artifact-freetrials needed. Based on such an input it skips themore artifactual EEG channels to replace singletrials (previously rejected for artifacts in one ormore channels) until the required number of trialsis reached. The software output is the list ofartifact-free trials and the electrode array to beused for further data analysis.

2.8. Recordings

Nine normal, right-handed young volunteersparticipated in the study. The general procedureswere approved from the local institutional ethicalcommittee, and all volunteers gave writteninformed consent. Each subject was seated in acomfortable reclining armchair placed in a dimlylit, electrically and acoustically shielded room.For both EEG and EOG, the pass-band and

sample rate were of 1–100 and 256 Hz, respec-tively. EEG data were recorded in 2 subjects) witha 128 electrodes cap and in seven subjects with a48 electrodes cap(Table 1). Linked earlobesserved as a reference. Electrode impedance waskept lower than 5–10 kV. Electrode positionswere measured with a sonic digitizer. Nasion,inion, and preauricular points were also digitizedfor subsequent integration between electrode posi-tions and MRI images.


Table 1The table shows the number of subjects, EEG, EOG or EMGtrials for each subject and number of recording electrodes

Subjects Artifacts No. of trials Electrodes

1 EEG–EOG–EMG 73 1282 VEOG 30 483 VEOG 23 484 VEOG 24 485 VEOG 23 486 VEOG 25 487 VEOG 25 48

8 No artifacts 50 128HEOG 25 48

9 VEOGqHEOG 25 48

VEOG, vertical EOG; HEOG, horizontal EOG.

Fig. 3. Flow chart of the algorithm for the channelytrial optimization. The software lists the numbers of the artifact free EEG singletrials. If the total number of trials is not high enough for the further data analysis, the software requires the indication of theminimum number of artifact-free trials as an input. Based on such an input, it skips the more artifactual EEG channels to replacetrials (previously discarded for artifacts) until the required number of trials is reached. The software output is the list of free-artifacttrials and the electrode array to be used for further data analysis. Symbols of the figure: Y, yes; N, no.

EOG activity was recorded with cup electrodesfor the control of blinking and eye movements. Acup electrode placed 1 cm above supra-orbitalridge registered the vertical EOG. It was referredto another electrode placed 2 cm below sub-orbitalridge of the right eye. The left and right horizontalEOG channels were collected from two electrodesat the left and the right lateral canthus. Theseelectrodes were referred to an electrode placed atthe glabella.EMG activity was recorded with surface cup

electrodes over bilateral extensor digitorum mus-cles(1–100 Hz pass-band; 256 Hz sampling rate;pair of Ag–AgCl electrodes positioned on botharms). This activity served to monitor hand muscleresponse of both sides across the experimentalconditions.


EEG recordings were carried out during a motortask in(2 subjects) and during a cognitive task(6subjects). The motor task consisted of a brisk self-paced right middle finger extensions followed bypassive return to the original resting position(inter-movement intervals of 6–20 s). Acquisition timewas of 8 s (from 4 s before to 4 s after thebeginning of the movement). On the other hand,the cognitive tasks consisted of a working memorytask with spatial-visuo-motor demand and its con-trol condition (no working memory). This taskwas a typical choice delayed reaction time task.Cue visual stimulus consisted of a couple ofvertical bars, followed by a few seconds delay.During the delay period, the subjects had toremember the position of the taller bar. After avisual-go stimulus, the subjects had to push leftbottom of mouse if the left vertical bar was tallerthan the right one or they had to push the rightbottom if the right vertical bar was taller than theleft one (the acquisition time was 12–14 s). Theno working memory task was like the workingmemory task, but the cue stimulus lasted to thevisual-go stimulus.Before the EEG recording, the subjects were

given 4–6 training sessions that establishedapproximately a stable level of motor performance.The total number of trials was of 323. The

number of trials for each subject is reported inTable 1.A control experiment was carried out to assess

if the standard threshold of 50mV is useful forthe recognition of both blinks and saccades and ifit is affected by small variations of the EOGelectrodes position. Other than the ‘standard’ ver-tical and horizontal EOG channels(see above),two additional EOG horizontal and vertical chan-nels were adopted. For the additional vertical EOGchannel, a ‘distant’ electrode was placed 1 cmabove the ‘standard’ electrode at right supra-orbitalridge. The additional horizontal channel used a‘distant’ electrode placed 1 cm lateral to the‘standard’ electrode at right outer canthus. These‘additional’ electrodes were referred to the sameelectrodes referencing the ‘standard’ electrodes.The subject performed visually triggered sac-

cades and blinks in the following blocks:(i) 50‘short’ and(ii) 59 ‘long’ horizontal saccades,(iii )

65 ‘short’ and (iv) 60 ‘long’ vertical saccades,and (v) 50 blinks. Based on visual cues appliedon the computer monitor, the ‘short’ saccades wereof 7.58 (horizontal) or 58 (vertical) whereas, the‘long’ saccades were of 158 (horizontal) or 108(vertical). The inter-trial interval of the visual cuewas of 4–8 s and the EEG acquisition time wasof 3 s (from 1 s preceding visual cue or trigger).

2.9. Waveforms and maps

To visualize the EEG waveforms of the EOGartifacts before and after their correction, 6 elec-trodes(Fp1, Fp2, C3, C4, P3, P4, representativeof the frontal, central and parietal regions of thescalp) were selected. The corresponding EEGpotentials were aligned and averaged with refer-ence to the peak of the averaged EOG artifacts.Averaged potentials distributions were visual-

ized with color topographical maps. These mapswere computed over a 3D(‘quasi-realistic’) brainmodel, by a regularized spline interpolating func-tion (Babiloni et al., 1998). This model is basedon the MRI images of 152 subjects, digitized atthe Brain Center of the Neurological Institute ofMontreal(SPM96).

2.10. Software validation

To test the software package, we selected 323trials from a large number of EOG–EMG–EEGtrials. The selected trials had been classified coher-ently (100%) by two skilled electroencephalo-graphists. The selected data set comprised 73artifactual trials for the detection of EMGresponses from the operating hand(onset, peak,offset of the EMG activity), of EMG artifacts fromthe non-operating hand(small EMG bursts due toinvoluntary mirror movements), and of EEG arti-facts due to instrumental and extra-cerebral causes.The instrumental EEG artifacts presented largepotential shifts provoked by bad electrode contactsand oscillations of the electrode cables. The extra-cerebral potentials showed rhythmic ‘electrocardi-ograph pulses’, EMG burst due to fronto-temporalmuscle activity, large potential shifts induced byswallowing and head movements. The selected


Table 2Analysis of concordancesydiscordances in the preliminaryanalysis of the EOG–EMG–EEG data between the selectionof the experimenters(gold standard) and software package

Mirror EOG EEG

N 49 250 51X 24.5 125 25.5n1 51 252 53n2 49 250 51F 1.843 1.309 1.8211a 0.01667 0.01667 0.01667L1 (%) 34.27 43.11 34.57L2 (%) 65.73 56.88 65.43

L1 and L2 values indicate statistical parameters of the inter-val out of which concordancesydiscordances between man andsoftware have a statistical significance(see text). Symbols:N,‘gold standard’ total trials number;Xs(0.5)=N; n1s2=(Xq1) (numerator degrees of freedom); n2s2=(NyX) (denom-inator degrees of freedom); a, test significance;F, value ofdistribution of F with significancea and degrees of freedomn1, n2.

Table 3Mean and standard deviation of the EMG onset, peak and offset values(ms) as measured by the experimenters(manually) andsoftware(automatically)

Automatic Manual Automatic Manual Automatic Manualonset onset peak peak offset offset

Total 73 73 73 73 73 73Mean 998.3"4.5 997.9"4.1 1033.1"4.7 1033.1"4.7 1067.7"6.8 1068.7"6.9

data set comprised also 200 trials with EOGartifacts. In particular, 150 trials presented verticalEOG artifacts, which are the most frequent artifactsin the EEG routine. Furthermore, 25 trials showedhorizontal EOG artifacts and vertical plus horizon-tal EOG artifacts characterized 25 trials. Finally,50 trials had no EOG–EMG–EEG artifacts.The analysis included the evaluation of discor-

danceyconcordance in the preliminary analysis ofEEG–EMG–EOG data between the selection ofthe experimenters(‘gold standard’) and softwarepackage(Table 2). In order to verify the statisticalsignificance of manysoftware concordance, a bino-mial test was used. The significance level(a) ofthe binomial test was of 0.01667(i.e. Bonferronicorrection foras0.05 and three repetitions of thetest for EEG, EMG, EOG data). The statistical

confidence limits were computed with the follow-ing formulas:

XL s1 UXq(NyXq1) FŽ .a,y1,y2

(Xq1)Fa,v1,v2L s2 UNyXq(Xq1) FŽ .a,v1,v2

3. Results

3.1. EMG artifact detection

The software package recognized approximately98% of the single trials with ‘mirror movements’artifacts, as detected by the two experimenters(gold standard).

3.2. EMG analysis (operant hand)

Table 3 shows the onset, peak and offset valuesof the EMG response as determined by the exper-imenters and software. There was a very highsimilarity of the EMG values accompanying vol-untary right finger movements, measured by exper-imenters(average of the experimenters’ choices)and software values. Consistently, there were nostatistical significant differences between thesevalues.

3.3. EOG artifacts

The software package recognized approximately95% of the single trials with artifacts in bothvertical and horizontal EOG channels(experi-menters’ classification as a gold standard). The


software amplitude threshold values of"40,"45,"50 and"55mV provided a similar rate of EOGartifact detection(94.82, 96.55, 96.55 and 94.82%,respectively). The highest percentage of recogni-tion (;97%) was observed with amplitude thresh-old values of"45 and"50 mV.For the control experiment, the mean("stan-

dard error) amplitudes of the EOG artifacts record-ed at ‘standard’ electrodes were ofy78.7"23.6(‘short’ horizontal saccades),y129.6"9.4(‘long’horizontal saccades), y71.08"19.5 (‘short’ ver-tical saccades), y86.8"32.02(‘long’ vertical sac-cades), y403.5"68.04 (blinks). Automaticartifact recognition was of 100% for the ‘short’and ‘long’ horizontal saccades, of 95.3–94.9% forthe ‘short’ and ‘long’ vertical saccades, and of100% for the blinks. Compared to the meanamplitudes of the ‘standard’ EOG electrodes, thoseof the ‘distant’ EOG electrodes were always lower(P-0.001). The mean amplitudes of the ‘distant’EOG electrodes were ofy53.8"8.3 (‘short’ hor-izontal saccades), y76.4"22.5 (‘long’ horizontalsaccades), y58.9"14.5 (‘short’ vertical sac-cades), y78.8"26.1 (‘long’ vertical saccades),andy331.6"62.1(blinks). The automatic artifactrecognition of the corresponding EOG artifactswas of 66%(‘short’) and 100%(‘long’) for thehorizontal saccades and of 73.8%(‘short’) and86.4% (‘long’) for the vertical saccades, and of100% for the blinks.After the correction of the EOG artifacts, the

percentage of the resumed EEG trials was asfollows: (i) 78.6% with OLS and 41.4% withFREQ for horizontal EOG artifacts;(ii) 74% withOLS, and 50% with FREQ for vertical EOGartifacts; (iii ) 69.6% with OLS, and 17.4% withFREQ for vertical plus horizontal EOG artifacts.Indeed, FREQ was less efficient than other meth-ods for the EOG artifacts correction.Fig. 4 shows EEG waveforms and topographical

maps before and after the vertical EOG correctionfor blinks by means of OLS, and FREQ methods.The EEG data were averaged with reference to thepeak of the EOG artifacts. The waveforms refer tosix representative electrode sites of 10–20 system.A similar effect of EOG artifact correction wasobserved at the frontal electrodes with all themethods. On the contrary, OLS showed the most

effective EOG artifact correction on the centraland posterior scalp electrode sites. With OLSmethod the residual signal is only of fewmV (11.5mV on average on the electrodes), fully in linewith the results obtained by other authors with‘regressive’ methods(Gasser et al., 1986; Kene-mans et al., 1991).In general, ‘regression’ methods in time(OLS)

and frequency(FREQ) domains performed aneffective correction of the ocular artifacts upon thefrontal electrodes Fp1 and Fp2. However, the moreeffective was executed with OLS compared toFREQ (P-0.05). Moreover, the EOG artifactswere corrected with OLS better than with FREQin the central and parietal electrodes sites(P-0.05 on C3, C4, P3, P4). On the whole, FREQresulted to be less efficient than OLS for the EOGartifacts correction and was not further consideredin this study.To verify that EOG artifact correction did not

introduce computational artifacts, OLS methodwas applied on a free-artifact data set(50 trials,subjects 8). A quantitative evaluation of thesecomputational artifacts was carried out by theestimation of statistical correlation. The spectralratio of EEG signals was then computed beforeand after the OLS correction. Negligible compu-tational artifacts would be correlated with highcorrelation values and low spectra ratio values.The correlation analysis confirmed this predic-

tion showing a high correlation, ranging from 0.8to 0.9 on the representative channels among EEGtraces before and after EOG artifact correction(Fig. 5). It should be noted that the presentcorrelation values have an indicative rather thanabsolute worth, given that a part of the correlationmight be ascribed to the effects of the referencesite used during the EEG recordings.Also, the spectral analysis did not show statis-

tical significant differences in the spectral featuresof the artifact-free EEG data before and after theuse of OLS method(Fig. 6). It can be seen thatOLS introduced no or negligible artifact aftercorrection. Of note, the spectral correlation wascomputed on the whole frequency range ratherthan only in the narrow frequency range where theocular artifact can be usually observed(i.e. deltaand teta). Indeed, there were no ocular artifacts in


Fig. 4. Effects of the EOG correction in the data set of single trials contaminated by ocular artifacts. To visualize the waveformsof the EOG artifacts, we selected six electrodes(Fp1, Fp2, C3, C4, P3, P4), which are representative of the frontal, central andparietal regions of the scalp. EEG potentials are aligned and averaged with reference to the onset time of the EOG artifact. Colormaps of potential amplitude at the instant of the peak of the EOG artifacts are computed over a 3D mean brain model(‘nearlyrealistic’). The maps show the peak amplitude of averaged scalp potentials before and after the correction of EOG artifacts witheach of the two proposed methods. Remarkably, OLS allows the best ocular artifacts correction at all considered electrodes. Therecordings have been retraced.

this data set and a main care of the analysis wasthe study of the possible distortions effects of OLSin frequency range of interest for the analysis ofcortical rhythmicity(i.e. alpha, beta, gamma).

3.4. EEG artifacts detection

The artifact detection of the EEG single trialsby experimenters(‘gold standard’) and software


Fig. 5. Effects of the EOG correction in the data set of singletrials (time domain) that are not contaminated by EOG arti-facts. In order to verify that ocular artifacts correction proce-dure did not introduce computational artifacts, OLS method isapplied on a free-artifact data set. The correlation among EEGtraces before and after correction is high. Such correlation is89%. The recordings have been retraced.

agreed at approximately 93%. Furthermore, theexperimenters and software concordance in rec-ognizing an artifact channels was of approximately96%.

4. Discussion

4.1. Methodological remarks on multi-centric EEGstudies

The impact of the reference electrode in EEGrecordings was beyond the scope of the presentstudy focused on preliminary EEG–EMG–EOGdata analysis. However, a brief discussion of thematter can be of interest(Nuwer, 1988). Unlikely,both cephalic and noncephalic references are not‘silent’ and contribute to the recorded EEG activity.A varying amount of EEG activity is present atcephalic reference sites and an increased probabil-ity of electrocardiographic and muscle activity canbe detected at noncephalic reference sites. Forexample, a frontal scalp reference could injectocular potential in all EEG channels and mastoidcan be prone to muscle activity(Pivik et al.,1993).For the present study, we used a very common

electrode reference for clinical and research pur-poses, i.e. the linked-earlobes. This allowed tomost readers a direct comparison of the presentEEG data(including EOG artifacts) with those of

their labs. As good points of linked-earlobe refer-ence, spontaneous generation of bioelectrical sig-nals is negligible at earlobes, the position is usuallynot tiresome for subjects, and that reference isquite distant from most cortical sources of EEG.Nevertheless, the physical linking of electrodescould provide a shunting of currents betweenelectrode sites that may distort the distribution ofthe scalp voltages(Miller et al., 1991). Further-more, the effect of unequal impedance in a linked-reference configuration would be to artificiallyinflate the amplitude of the leads on the side ofthe reference electrode of highest resistance(Gar-neski and Steelman, 1958). Even if the equaliza-tion can be achieved at the onset of a recordingsession, these values may change over the courseof that session.A first strategy for a correct EEG recording with

linked earlobe is to place a resistor in series witheach electrode prior to electrically linking theleads. A second strategy is to place a variableresistor in series with each active electrode, con-stantly monitor electrode impedances, and balancethe impedances or changes in impedance over thecourse of experiment. A third strategy is to dedi-cate separate preamplifier channel to each refer-ence electrode and to electrically link the outputsof two amplifiers (Pivik et al., 1993). Previousstudies have indicated that these procedures canprovide reliable EEG data(Senulis and Davidson,1989; Andino et al., 1990). However, the use oflinked ear reference should be made with cautious,at least when the impedances of the two earlobesare not properly controlled across the recordingsession.In general, a reliable way to manage the refer-

ence issue is the recording of EEG data using asingle(i.e. cephalic) reference and the subsequentcomputation of the so called ‘common average’(Picton et al., 2000a), i.e. an averaging calculatedas the sum of the EEG activity in all recordedchannels divided by the number of electrodes plusone (Dien, 1998; Picton et al., 2000a). Commonaverage reference is particularly suitable for spatialanalysis of EEG data including source analysis(i.e. EEG data not biased by a sample referencesite) and for correlation-based analysis(correla-tions not inflated by the activity at a single


Fig. 6. Effects of the EOG correction in the data set of single trials(frequency domain) that are not contaminated by EOG artifacts.In order to verify that ocular artifacts correction procedure did not introduce computational artifacts, we performed a spectral analysisbefore and after correction with OLS. The spectral analysis does not show statistical significant differences among the spectra beforeand after the EOG correction. In fact, differences are maintained within the confidence interval(P-0.05).

reference site). In particular, the common averagereference is optimal for the detection of electrodeartifacts, because it maximizes the effects of eachindividual electrode at each data channel(Pictonet al., 2000a). It is a very valid alternative to off-line surface Laplacian estimation(Babiloni et al.,2001) at least for making reference-free the EEGdata.

4.2. EOG detection and correction

In the present study, the automatic EOG artifactdetection made by software was very similar tothat of the experimenters(;95% of concordance).

It is important to underline that the two skilledexperimenters selected the trials with ocular arti-facts in a totally subjective way, while the softwarebased its selection upon a simple 50-mV rejectionthreshold. In particular, the two experimentersperformed the preliminary data analysis based onthe following criteria: (1) substantial magnitudeof EEG artifacts respect to background EEG oscil-lations; (2) typical bell-like shape of fast EOGartifacts(blinks); and (3) typical ramp-like shapeof horizontal or vertical EOG artifacts due tosaccades.The variation of the detection threshold from

"40 to"55 mV introduced only little changes in


the percentage(;2%) of the manysoftwareconcordance.The results of the present study indicated that

with EOG electrodes placed in the ‘standard’position (Section 2), the threshold of 50mV isreliable for the automatic detection of not onlyblinks but also of vertical and horizontal saccades.In contrast, EOG electrode positions different ofonly 1 cm from ‘standard’ electrodes stronglyreduced the EOG mean amplitude(P-0.0001)and the percentage(;30%) of the artifact recog-nition for both ‘short’ vertical and horizontalsaccades(5–7.58). The EOG electrodes shouldthen be placed with particular care by the operativeunits of a multi-centric EEG study. A simplecontrol of this variable could be performed by theregistration of the electrodes position by magneticdigitizer or web cam.In the present study, EOG artifact correction

was carried out by two ‘regression’ methods oper-ating in both time(i.e. OLS) and frequency(i.e.FREQ) domains. These methods were chosenbecause they are well known in literature, reliable,and suitable at a first glance for EEG multi-centricstudies (Whitton et al., 1978; Fortgens and deBruin, 1983; Kenemans et al., 1991; Wauschkunet al., 1998; Picton et al., 2000a). Noteworthy, thefrequency domain method would have performedwell using a transfer function, averaged by severalsubjects for each band of frequency(Gasser et al.,1986). This procedure would have made the meth-od more balanced. However, it is not very easy touse in the framework of a multi-centric EEG study.The present results showed that a better verticaland horizontal EOG correction was obtained withOLS than FREQ. Plausibly, this is due to the factthat OLS computed the parameters for the EOGcorrection segment-by-segment of EEG trials rath-er than on the whole EEG trial.The present results extended previous evidence

that ‘regression’ methods perform poorly whenblinks and eye movements occur too frequently intime (Picton et al., 2000a). We observed that theEOG correction by these methods was globallymuch better with artifacts occurring at only oneEOG channel(vertical or horizontal) rather thanat both EOG channels. The poor performance ofthe ‘regression’ methods would be due to the fact

that simultaneous vertical and horizontal EOGartifacts are associated with further sources ofartifacts increasing the level of noise in the EEGdata. Unlikely, the ‘regression’ methods are notable to separate in topography different sources ofEEG artifacts, when compared to EOG correctionmethods based on PCA or ICA or dipole modeling(Berg and Scherg, 1991; Vigario, 1997; Jung etal., 1999, 2000; Picton et al., 2000b).Another theoretical weakness of the ‘regression’

methods is that they could in part remove thesignal component propagated from EEG to EOGleads. Our findings showed that such a signalcancellation is practically negligible at the level ofEEG waveforms in both time and frequencydomains, at least on the present data set in whichthere were no EOG artifacts. Indeed the use of theOLS method induced no computational artifacts inthe corrected EEG data. However, EOG ‘regres-sion’ correction methods(i.e. especially in fre-quency domain) might change the topographicdistribution of the EEG(Picton et al., 2000b) andthen the results of procedures aimed at localizingbrain sources of EEG data(i.e. surface Laplacianestimation; Babiloni et al., 1996). Therefore, weneed further studies to delineate well the extent towhich EOG correction methods distort brain sig-nals in the process of removing EOG artifacts.In the present study, we did not consider EOG

correction methods based on PCA or ICA or dipolemodeling (i.e. ‘spatial components’ methods),which are conceived to separate in topography andcorrect vertical and horizontal EOG as well asEEG artifacts(Berg and Scherg, 1991; Lagerlundet al., 1997; Jung et al., 1999, 2000). As afore-mentioned, these methods need much more expertcontrol by experimenters during the setting of theEOG correction parameters(PCA and ICA) or anexplicit mathematical modeling of eyes as dipolesources and head as a volume conductor. The‘spatial components’ methods should be alwayspreferred to ‘regression’ methods when either ver-tical or horizontal EOG channels are not availablefor the artifact correction. Indeed, the effective useof the ‘regression’ methods strictly depend on theavailability of vertical and horizontal EOG chan-nels. The accuracy of the ‘spatial components’method based on the PCA depends on the spatial


separation and difference in amplitude of thesources of EEG data and EOG artifacts as well ason the orthogonality of these sources(Jung et al.,1999, 2000). These limitations can be overcomeby the spatial component method based on theICA (Jung et al., 1998, 2000) or dipole modelingmethod(Berg and Scherg, 1991).

4.3. EMG data analysis

In our experiments, the percentage of concor-dance among the choices of the experimenters andsoftware in the detection of EMG artifacts fromthe non-operant hand(‘mirror movements’) wasvery high (97.97%). Also, the analysis of move-ment-related EMG response(onset, peak and offsetvalues) in the operant hand showed a similarconcordance between the experimenters(‘goldstandard’) and software. The proposed softwarerecognized the onset and offset of EMG, based ontwo steps. In the first step, the software sets threeamplitude thresholds related to the backgroundnoise (absolute baseline) and to the partial andcomplete synchronization of the motor unitsinvolved in the motor performances, respectively.The EMG peak was chosen as a reference for theassignment of the second and third amplitudethresholds. In fact, the EMG peak is related to thegreatest action potentials synchronization of motorunits during an active muscular contraction(Arri-go, 1991; Pinelli, 1993). In general, the firstamplitude threshold exceeded the backgroundEMG noise. In agreement with previous study(Leader et al., 1998; Micera et al., 1998), such athreshold is equivalent to EMG activity averagedin the first second of signal plus three standarddeviations. The procedure seems to be efficacious,since EMG signal-to-noise ratio can be reasonablysupposed to be elevated(Latwesen and Patterson,1994; Micera et al., 1998).The other two amplitude thresholds divided the

EMG signal in two identical parts, correspondingto the 45 and 90% of EMG peak amplitude.Variations up to 5–10% of such thresholds werenot influent on the algorithm performances, due tothe intrinsic stability of EMG signal during briskvoluntary movements. In such a condition, a highsynchronization of the motor units can be

observed. That is why the resulting EMG responseis constant among different trials and subjects(Finucane et al., 1998; Kollmitzer et al., 1999).Coherently, previous studies have shown an ele-vated reliability of surface EMG parameters withina maximal or sub-maximal muscular contraction(Kollmitzer et al., 1999). In the second step, thesoftware computes four short EMG periods in thetemporal sequence(lasting 15-ms each). The firstperiod served as a reference period(relative base-line) for the following three periods. The EMGonset time was defined as the instant after whichamplitude values in the three subsequent periodsexceed three standard deviations of the relativebaseline period. In the same time, the signal hadto exceed all the amplitude thresholds(‘noise’, 45and 90% of the EMG peak). It is important tonote that our method is a novel variant of classicmethods for the detection of the onset in EMGresponse. Algorithms like generalized likelihoodratio (GLR) would have been preferable in pres-ence of a low signal-to-noise-ratio, but not in ourcondition (Micera et al., 1998). Indeed, move-ment-related EMG responses present a high signal-to-noise-ratio. This is due to the fact that GLRdoes not model the real EMG signal according tothe gaussian features(Latwesen and Patterson,1994). Moreover, our software presents a temporalstep, which works together with a quantitativestep. This check is present in the detection of bothEMG onset and offset.

4.4. EEG data analysis

The most common methods to insulate the EEGnoise are essentially three:(1) rejection threshold-based methods(2) detection statistical-based meth-ods and (3) removal filtering-based methods(Cutmore and Daniel, 1999). Our software per-formed its detection on both a statistical criterion,based on EEG signal non-stationarity and ampli-tude threshold method(default"100mV).The concordance among the choices of the

experimenters and software was very high(93%)in the detection of the EEG artifacts. The EEGamplitude threshold can be modified according tothe features of subjects involved in the study. Forexample, it can be maintained at"100 mV in


dementia studies(Adams et al., 1998; Samuels,1998) and elevated over"100 mV in subjectswith epilepsy or metabolic encephalopathy(Adams et al., 1998; Samuels, 1998).With the non-stationarity criterion, the mean

EEG amplitudes of two contiguous 500-ms seg-ments was supposed to be statistically stationaryat each EEG channel. This condition may be trueeven across an EEG trial related to cognitive–sensorimotor events, since we consider large bandEEG changes rather than narrow band EEG chang-es. Indeed, the event-related potential is approxi-mately 10 times lower in amplitude than thebackground EEG oscillations, so that it does notchange the statistical stationarity of contiguousEEG segments. Furthermore, the cognitive or sen-sorimotor events can induce changes in EEGoscillations, but only in specific frequency bands,i.e. the rapid post-event enhancement of alpha(8–12 Hz) or beta (16–22 Hz) band (Pfurtscheller,1988; Basar et al., 2001). The event-related chang-es of EEG oscillations at narrow frequency bandswould not result in wrong EEG artifact identifica-tion, based on the ‘stationarity’ criterion. In fact,the software package showed statistical stationarityin our artifact-free EEG data related to sensori-motor events. Instead, the software pointed toapproximately 95% non-stationarity in the periodscorresponding to the artifacts detected by theexperimenters. It is noteworthy that all EEG singletrials of the present study were recorded duringboth sensorimotor and cognitive events. Therefore,these events do not induce ‘per se’ non-stationaritywhen the artifact detection is performed in thelarge band EEG signals. Of note is that mathemat-ical approaches for EEG artifact detection orremoval such as filtering-based methods were nottaken into account here, since they are not suitablefor multi-centric EEG studies(i.e. need for reduc-tion of manual revision and data processing byexpert electroencephalographists, etc.).A final optimization procedure of the software

package suggests an artifact-free trialsychannelsratio for further analysis. The experimenter canpreliminarily indicate the minimum number ofartifact-free EEG trials. If the actual artifact-freeEEG trials are insufficient to fit that number, thesoftware replaces trials, rejecting from the elec-

trode montage the most artifactual channels. Thisextends standard algorithms for removing badelectrodes when there are few trials in the average.

5. Conclusions

In the present study, we presented a softwarepackage for a fast and automatic selection of EEGdata free from EOG, EMG and EEG artifacts. Thesoftware collection integrated inMATLAB environ-ment are well known criteria(i.e. amplitude thresh-old) and algorithms(i.e. EOG artifact correction)are already accepted by the scientific community.As a main contribution of this study, the softwarepackage was validated on a large EOG–EMG–EEG data set using the classification of two expertexperimenters as a ‘gold standard’. A specialattention was devoted to compare the performancesof two well known ‘regression’ methods for theEOG artifacts correction.The results showed an acceptable agreement

(;95%) between the human and software behav-iors for (i) the detection of vertical and horizontalEOG artifacts,(ii) the measurement of hand EMGresponses for a cognitive–motor paradigm,(iii )the detection of involuntary mirror movements,and (iv) the detection of EEG artifacts. Further-more, our results indicated a particular reliabilityof a ‘regression’ EOG correction method operatingin time domain(i.e. OLS). These results suggestthat such a software package could be used formulti-centric EEG studies and, indeed, we use itin the project ‘Alzheimer Disease EEG Databaseon-line’ (http:__hreeg.ifu.uniroma1.it).Finally, it should be make clear that, ideally,

this software package should be used for an initialscreening of the data of a multi-centric EEG study,to be followed by a final manual revision per-formed by an expert electroencephalographist. Thehuman ‘eyes’ are still the most reliable ‘softwarepackage’ to date. A totally automatic use of thesoftware package can then be accepted only whenthe amount of data from multi-centric EEG studiesis incompatible with the work of a single researchunit.


Acknowledgments

The authors would like to thank Prof. FabrizioEusebi, Chairman of the Biophysics Group ofInterest of the Department of Human Physiologyand Pharmacology—University of Rome ‘La Sap-ienza’—for his continuous support. The researchwas granted by Italian Ministry of University,Fatebenefratelli Association for Research(AFaR),and Foundation Telethon Onlus(Project E.C0985).

References

Adams, R.D., Victor, M., Hopper, A.H.(Eds.), 1998. Princi-ples of Neurology. McGraw-Hill, New York, pp. 1998.

Andino, F.L.G., Marqui, R.D.P., Sossa, P.A.V., et al., 1990.Brain electrical field measurements unaffected by linkedearlobes reference. Electroencephalogr. Clin. Neurophysiol.75, 155–160.

Arrigo, M., Carreras, D., De Grandis, D., Serra, G.(Eds.),1991. Elementi di Elettromiografia e Neurofisiologia Clini-ca. Editore Marrapese, Roma, pp. 1991.

Aurlien, H., Gjerde, I.O., Gilhus, N.E., 1999. A new way ofbuilding a database of EEG findings. Clin. Neurophysiol.110, 986–995.

Babiloni, C., Babiloni, F., Carducci, F., et al., 2000. Movement-related electroencephalographic reactivity in Alzheimer dis-ease. Neuroimage 12, 139–146.

Babiloni, C., Carducci, F., Cincotti, F., et al., 1999. Humanmovement-related potentials vs. desynchronization of EEGalpha rhythm: a high-resolution EEG study. Neuroimage 10,658–665.

Babiloni, F., Babiloni, C., Carducci, F., Fattorini, L., Onorati,P., Urbano, A., 1996. Spline Laplacian estimate of EEGpotentials over a realistic magnetic resonance-constructedscalp surface model. Electroencephalogr. Clin. Neurophy-siol. 98, 363–373.

Babiloni, F., Carducci, F., Babiloni, C., Urbano, A., 1998.Improved realistic Laplacian estimate of highly-sampledEEG potentials with regularization techniques. Electroence-phalogr. Clin. Neurophysiol. 106, 336–343.

Babiloni, F., Cincotti, F., Carducci, F., Rossini, P.M., Babiloni,C., 2001. Spatial enhancement of EEG data by surfaceLaplacian estimation: the use of magnetic resonance imag-ing-based head models. Clin. Neurophysiol. 112, 724–727.

Basar, E., Basar-Eroglu, C., Karakas, S., Schurmann, M., 2001.Gamma, alpha, delta, and theta oscillations govern cognitiveprocess. Int. J. Psychophysiol. 39, 241–248.

Bell, A.J., Sejnowski, T.J., 1995. An information–maximiza-tion approach to blind separation and blind deconvolution.Neural Comput. 7, 1129–1159.

Berg, P., Scherg, M., 1991. Dipole models of eye movementand blinks. Electroencephalogr. Clin. Neurophysiol. 79,36–44.

Berg, P., Scherg, M., 1994. A multiple source approach to thecorrection of eye artifacts. Electroencephalogr. Clin. Neu-rophysiol. 90, 229–241.

Croft, R.J., Barry, R.J., 1998a. EOG correction: a new per-spective. Electroencephalogr. Clin. Neurophysiol. 107,387–394.

Croft, R.J., Barry, R.J., 1998b. EOG correction: a new aligned-artifact average solution. Electroencephalogr. Clin. Neuro-physiol. 107, 395–401.

Croft, R.J., Barry, R.J., 2000a. Removal of ocular artifact fromthe EEG: a review. Clin. Neurophysiol. 49, 5–19.

Croft, R.J., Barry, R.J., 2000b. EOG correction of blinks withsaccade coefficients: a test and revision of aligned-artifactaverage solution. Clin. Neurophysiol. 111, 444–451.

Croft, R.J., Barry, R.J., 2000c. EOG correction: which regres-sion should we use? Psychophysiology 37, 123–125.

Cutmore, T.R.H., Daniel, A.J., 1999. Identifying and reducingnoise in psychophysiological recordings. Int. J. Psychophy-siol. 32, 129–150.

Daly, D.D., Pedley, T.A.(Eds.), 1990. Current Practice ofClinical EEG. Raven Press, New York, pp. 1990.

Dien, J., 1998. Issues in the application of the averagereference: review, critiques, and recommendations. Behav.Res. Methods Instruments Comput. 30, 34–43.

Finucane, S.D.G., Rafeei, T., Kues, J., et al., 1998. Reproduc-ibility of electromyographic recordings of submaximal con-centric and eccentric muscle contractions in humans.Electroencephalogr. Clin. Neurophysiol. 109, 290–296.

Fortgens, C., de Bruin, M.P., 1983. Removal of eye movementand ECG artifacts from the non-cephalic reference EEG.Electroencephalogr. Clin. Neurophysiol. 56, 90–96.

Garneski, T.M., Steelman, H.F., 1958. Equalizing ear referenceresistance in monopolar recording to eliminate artifactualtemporal lobe asymmetry. Electroencephalogr. Clin. Neuro-physiol. 10, 736–737.

Gasser, T., Sroka, L., Mocks, J., 1986. The correction of EOGartifacts by frequency dependent and frequency independentmethods. Psychophysiology 23, 704–712.

Gratton, G., Coles, M.G.H., Donchin, E., 1983. A new methodfor the off-line removal of ocular artifact. Electroencephal-ogr. Clin. Neurophysiol. 55, 468–484.

Hillyard, S.A., Galambos, R., 1970. Eye-movement artifact inthe CNV. Electroencephalogr. Clin. Neurophysiol. 28,173–182.

Hughes, J.R.(Ed.), 1994. EEG in Clinical Practice. Butter-worth, Woburn, Massachusetts, pp. 1994.

Jerwis, B.W., Nichols, M.J., Allen, E.M., Hudson, N.R.,Johnson, T.E., 1985. The assessment of two methods forremoving eye movement artifact from the EEG. Electroen-cephalogr. Clin. Neurophysiol. 61, 444–452.

Jung, T.-P., Humphries, C., Lee, T.-W., et al., 1998. Removingelectroencephalographic artifacts: comparison between ICAand PCA. Neural Networks Signal Process. 8, 63–72.


Jung, T.-P., Makeig, S., Westerfield, M., Townsend, J., Cour-chesne, E., Sejnowski, T.J., 1999. Analyzing and visualizingsingle-trial event-related potentials. Adv. Neural Info. Pro-cess. Systems 11, 118–124.

Jung, T.-P., Makeig, S., Westerfield, M., Townsend, J., Cour-chesne, E., Sejnowski, T.J., 2000. Removal of eye activityartifacts from visual event-related potentials in normal andclinical subjects. Clin. Neurophysiol. 111, 1745–1758.

Kenemans, J.L., Molenaar, P.C.M., Verbaten, M.N., Slangen,J.L., 1991. Removal of the ocular artifact from the EEG: acomparison of time and frequency methods with simulatedand real data. Psychophysiology 28, 114–121.

Kollmitzer, J., Ebebinchler, G.R., Kopf, A., 1999. Reliabilityof surface electromyographic measurements. Clin. Neuro-physiol. 110, 725–734.

Lagerlund, T.D., Sharbrough, F.W., Busacker, N.E., 1997.Spatial filtering of multichannel electroencephalographicrecordings through principal component analysis by singularvalue decomposition. J. Clin. Neurophysiol. 14, 73–82.

Latwesen, A., Patterson, P.E., 1994. Identification of lowerarm motions using the EMG signals of shoulder muscles.Med. Eng. Phys. 16, 113–121.

Leader, J.K., Boston, J.R., Moore, C.A., 1998. A data depend-ent computer algorithm for the detection of muscle activityonset and offset from EMG recordings. Electroencephalogr.Clin. Neurophysiol. 109, 119–123.

Leuchter, A.F., Cook, I.A., Newton, T.F., et al., 1993. Regionaldifferences in brain electrical activity in dementia: use ofspectral power and spectral ratio measures. Electroence-phalogr. Clin. Neurophysiol. 87, 385–393.

Lins, O.G., Picton, T.W., Berg, P., Scherg, M., 1993a. Ocularartifacts in EEG and event-related potentials I. Scalp topog-raphy. Brain Topogr. 6, 51–63.

Lins, O.G., Picton, T.W., Berg, P., Scherg, M., 1993b. Ocularartifacts in EEG and event-related potentials. II Sourcedipole and source components. Brain Topogr. 6, 65–78.

Locatelli, T., Cursi, M., Liberati, D., Franceschi, M., Comi,G., 1998. EEG coherence in Alzheimer’s disease. Electroen-cephalogr. Clin. Neurophysiol. 106, 229–237.

Makeig, S., Westerfield, M., Townsend, J., Jung, T.-P., Cour-chesne, E., Sejnowski, T.J., 1999. Independent componentanalysis of early event-related potentials in a visual spatialattention task. Phil. Trans. Biol. Sci. 354, 1135–1144.

Micera, S., Sabatini, A.M., Dario, P., 1998. An algorithm fordetecting the onset of muscle contraction by EMG signalprocessing. Med. Eng. Phys. 20, 211–215.

Miller, G.A., Lutzenberger, W, Elbert, T., 1991. The linked-reference issue in EEG and ERP recording. J. Psychophysiol.5, 273–276.

Nuwer, M.R., 1988. Quantitative EEGI. Techniques and prob-lems of frequency analysis and topographic mapping. J.Clin. Neurophys. 5, 1–43.

Neubauer, A.C., Sange, G., Pfurtscheller, G., 1999. Psycho-metric intelligence and event-related desynchronization dur-

ing performance of a letter matching task. In: Pfurtscheller,G., Lopes da Silva, F.H.(Eds.), Event-Related Desynchron-ization Handbook of Electroencephalography and ClinicalNeurophysiology. Elsevier, Amsterdam.

Picton, T.W., Bentin, S., Berg, P., et al., 2000a. Guidelines forusing human event-related potentials to study cognition:recording standards and publication criteria. Psychophysi-ology 37, 127–152.

Picton, T.W., van Roon, P., Armilio, M.L., Berg, P., Ille, N.,Scherg, M., 2000b. The correction of ocular artifacts: atopographic perspective. Clin. Neurophysiol. 111, 53–65.

Pinelli, P., 1993. In: Pinelli, P., Poloni, M.(Eds.), Neurologia:Principi di Diagnostica e Terapia. Casa Editrice Ambrosiana,Milano, pp. 1993.

Pivik, R.T., Broughton, R.J., Coppola, R., Davidson, R.J., Fox,N., Nuwer, M.R., 1993. Guidelines for the recording andquantitative analysis of electroencephalographic activity inresearch context. Psychophysiology 30, 547–558.

Pfurtscheller, G., 1988. In: Pfurtscheller, G., Lopes da Silva,F. (Eds.), Functional Brain Imaging. Hans Huber, pp. 1988.

Pfurtscheller, G., Lopez da Silva, F., 1999. Event-related EEGyMEG synchronization and desynchronization: basic princi-ples. Clin. Neurophysiol. 110, 1842–1857.

Samuels, M., 1998. In: Samuels, M.A., Feske, S.(Eds.),Office Practice of Neurology. Churchill Livingstone, NewYork.

Senulis, J.A., Davidson, R.J., 1989. The effect of linking theears on the hemispheric asymmetry of EEG. Psychophysi-ology 26, S54.

Urbano, A., Babiloni, C., Onorati, P., Babiloni, F., 1996.Human cortical activity related to unilateral movements. Ahigh resolution EEG study. Neuroreport 8, 203–206.

Van den Berg-Lenssen, M.M.C., Van Gisbergen, J.A.M., Jervis,B.W., 1994. Comparison of two methods for correctingocular artifacts in EEGs. Med. Biol. Eng. Comput. 32,501–511.

Verleger, R., 1993. Valid identification of blinks artifacts: arethey larger than 50mV in EEG records? Electroencephalogr.Clin. Neurophysiol. 87, 354–363.

Verleger, R., Gasser, T., Mocks, J., 1982. Correction of EOGartifacts in event related potentials of the EEG: aspects ofreliability and validity. Psychophysiology 19, 472–480.

Vigario, R.N., 1997. Extraction of ocular artifacts from EEGusing independent component analysis. Electroencephalogr.Clin. Neurophysiol. 103, 395–404.

Wauschkun, B., Verleger, R., Wascher, E., et al., 1998. Later-alized human cortical activity for shifting visuospatial atten-tion and initiating saccades. J. Neurophysiol. 80, 2900–2910.

Whitton, J.L., Lue, F., Moldofsky, H., 1978. Aspectral methodfor removing eye-movement artifacts from the EEG. Elec-troencephalogr. Clin. Neurophysiol. 44, 735–741.

Woestenburg, J.C., Verbaten, M.N., Slangen, J.L., 1983. Theremoval of the eye-movement artifact from the EEG byregression analysis in the frequency domain. Biol. Psychol.16, 127–147.

Date post:	01-Feb-2017
Category:	Documents
Upload:	nguyenkhanh
View:	218 times
Download:	0 times

D.V. Moretti et al., 2003

Documents