Download - Using Eye Tracker for Accurate Eye Movement Artifact Correction

1256 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 7, JULY 2007

Using an Eye Tracker for Accurate Eye MovementArtifact Correction

Joep J. M. Kierkels*, Jamal Riani, Jan W. M. Bergmans, Senior Member, IEEE, and Geert J. M. van Boxtel

AbstractWe present a new method to correct eye movementartifacts in electroencephalogram (EEG) data. By using an eyetracker, whose data cannot be corrupted by any electrophysio-logical signals, an accurate method for correction is developed.The eye-tracker data is used in a Kalman filter to estimate whichpart of the EEG is of ocular origin. The main assumptions foroptimal correction are summed and their validity is proven. Theeye-tracker-based correction method is objectively evaluated onsimulated data of four different types of eye movements and visu-ally evaluated on experimental data. Results are compared to threeestablished correction methods: Regression, Principal ComponentAnalysis, and Second-Order Blind Identification. A comparisonof signal to noise ratio after correction by these methods is givenin Table II and shows that our method is consistently superiorto the other three methods, often by a large margin. The useof a reference signal without electrophysiological influences, asprovided by an eye tracker, is essential to achieve optimal eyemovement artifact removal.

Index TermsArtifact removal, electroencephalography, eyemovements, modeling.

I. INTRODUCTION

TO CORRECT the electroencephalogram (EEG) for eyemovement and blink artifacts, many correction methodshave been developed over the past years [1][4]. Especiallyin research areas where the EEG signals of interest have verylow amplitudes and are of short duration, as for single-trialexperiments, it is important that the correction method removesas much of the artifact as possible. Often, like in habituationstudies or in studies involving children or ADHD subjects, itis not possible or undesired to repeat the experiment numeroustimes if artifacts occur. Furthermore, the electrical activity ofbrain processes that mainly occur in the frontal lobe is difficultto detect because frontal electrode positions can contain eyemovement artifacts of large amplitude.

Both brain activity and eye movements cause electric cur-rents through the brain. Therefore, the recorded EEG signal isa combination of ocular and brain-related components. Aftera recorded signal is corrected for ocular artifacts (OAs) it is

Manuscript received January 1, 2006; revised October 29, 2006. This workwas supported by a grant from the Co-operation Centre Tilburg and EindhovenUniversities. Asterisk indicates corresponding author.

*J. J. M. Kierkels is with the Electrical Engineering Department, EindhovenUniversity of Technology, P.O. Box 513, 5600 MB, Eindhoven, The Nether-lands (e-mail: [email protected]).

J. Riani and J. W. M. Bergmans are with the Electrical Engineering Depart-ment, Eindhoven University of Technology, 5600 MB, Eindhoven, The Nether-lands.

G. J. M. van Boxtel is with the Psychology Department, University of Tilburg,5000 LE, Tilburg, The Netherlands.

Digital Object Identifier 10.1109/TBME.2006.889179

difficult to judge if, and to what extent, correction was suc-cessful because the brain and ocular components are not sep-arately known. For this reason, it is also not yet possible to ob-jectively determine the quality of existing correction methods,and hence their adequacy for challenging applications like thosementioned above.

In order to develop a standard against which existing methodscan be compared, it is necessary to have a method that, in prin-ciple, can achieve optimal correction. The goal of this study isto develop such a method and use it to objectively determine thequality of correction of existing methods.

All existing correction methods are, to our knowledge, purelybased on electrical potential recordings. If either the ocular orthe brain component in the EEG can be reconstructed withoutthe other, it is possible to extract both components from themixture and objectively determine the quality and adequacy ofcorrection methods. The ocular component is caused by a dif-ference in potential between the front and the back of the eye,known as the corneo-retinal dipole [5]. Eye movements changethe orientation of this dipole and thus, via volume conductionthrough the head, also change the magnitude of the ocular com-ponent. Eye blinks and smaller eyelid movements also causechanges in potential at the electrode positions, which can resultin artifacts with amplitudes of up to 300 . The origin of thechange in potential as caused by blinks is different from eyemovement potential changes. Blinks briefly change the shapeof the volume that surrounds the corneo-retinal dipole. As aresult, the attenuation of blink artifacts from frontal to occip-ital electrodes is different from the attenuation of eye move-ment artifacts. Moreover, the specific influences of eye move-ments or eyelid movement on the EEG are difficult to discern.Many studies have demonstrated that there is an accompanyingeye movement during a blink and, similarly, during most eyemovements there is an accompanying eyelid movement [6][8].Modeling these two artifacts requires two different approaches.In this paper, the focus is on both simulated and recorded eyemovement artifacts. By omitting the effects of blinks and eyelidmovement during eye movements in our simulations, a consid-erable simplification is made. A correction method that claimsto correct for both blinks and for eye movements should, how-ever, be able to correct the data presented here as well, becausethe eyelid position is fixed in our simulations.

We propose to record the orientation of the eye by an eyetracker in order to provide information on the OA that doesnot contain any cerebral component. As a measure that repre-sents the orientation of the dipole, the eye-tracker records thehorizontal and vertical position of the pupil, denoted byand , respectively. These positions, combined in a vector

, are indicative of the ocular orientation.

0018-9294/$25.00 2007 IEEE

KIERKELS et al.: USING AN EYE TRACKER FOR ACCURATE EYE MOVEMENT ARTIFACT CORRECTION 1257

Fig. 1. Use of an eye tracker as a basis for eye movement artifact correction.The enclosed shape in this figure represent the human part of the setup. Solidarrows indicate potentials.

In Fig. 1, it is illustrated that a recorded EEG, , containspotentials of both cerebral, , and ocular origins, with

for eye. The potential is determined by the ocular ori-entation and by the conductive properties of the head and is as-sumed to be a function of . The separation of the compo-nents in is illustrated in the lower part of Fig. 1. Becausechanges in have instantaneous effects on electric potential,due to volume conduction through the head [9], vector canbe converted to an estimate of , denoted by . Forthis conversion, it is necessary that the conductive properties ofthe head are parameterized in a way that allows for the calcula-tion of based on the vector . By subtractingfrom , an estimate for the cerebral component can also beobtained, denoted as .

The relation between and is unknown and de-pends, among other things, on physical properties of the subject,like the diameter of the head and exact morphology of the skull,brain and other biological tissues. Obviously, the relation alsodepends on nonsubject-related properties, like electrode place-ment and the luminance over the retina.

In this paper, it is assumed, and verified, that this relationcan be parameterized by using first- and second-order combina-tions of . The resulting parameters, combined in a parametervector , are a priori unknown as they represent the physical-and nonsubject-related properties discussed earlier, and must beestimated based on recorded data. For accurate artifact removal,it is essential that this estimation is accurate.

Traditionally, artifact correction also combined several un-known elements in one parameter or a vector of parame-ters. The main difference with the current approach is thatonly focuses on the relation between pupil position and recordedEEG.

The vector is usually estimated nonadaptively, eitherduring a calibration session, or directly on the data of interest,and leading to estimate vector . Nonadaptive methods esti-mate a constant , over a period of time. Fluctuations ofin time will result in sub-optimal correction as a fluctuation ofonly one percent can cause new artifacts of several .

This can be overcome in two ways. Firstly, the length of therecording can be reduced to decrease fluctuations of withinthis recording. Examples of such are recordings during which

is recalibrated at fixed times, or component analysis over anepoch of only a few seconds. The effects of a shorter epoch onaccuracy of regression-based and component-based methods isstudied, e.g., in [10].

In this study, it was found that small parameter fluctuationsare less problematic for regression-based methods. Furthermorecorrection over a 60 s epoch was significantly worse than cor-rection over a 1 s epoch, which supports the idea that doesfluctuate. A difficulty with this approach is that after correctionthe epochs need to be re-attached and jumps may occur. Sec-ondly, by using parameter adaptation to track , it is pos-sible to adapt to parameter fluctuations and have an accurate

throughout a recording of any duration. This results in asmooth corrected signal that does not suffer from re-attachingproblems. Given these advantages, adaptive parameter estima-tion is used in this study.

Vector is, thus, obtained adaptively by a feedback-loopin which is used as a basis for adaptation. Adaptation isindicated in Fig. 1 by the dashed gray arrow. The motivationbehind the use of at most second-order parameterization is that

is expected to have only one extremum occuring if theocular dipole is oriented towards the electrode. Moving awayfrom this orientation, the changes in potential are expected tobe smooth.

For the method to work, it is necessary that three requirementsare met.

1) Changes in ocular orientation have instantaneous effects onelectric potential.

2) Ocular orientation is statistically independent of brain ac-tivity, as reflected in the EEG.

3) The relation between and is adequately pa-rameterized by first- and second-order combinations of

, i.e., higher-order terms are negligible.Assuming that these requirements are met, we will demonstratethat optimal correction can, in principle, be achieved, such thatno OA remains. Requirement one, i.e., instantaneous conduc-tion, has been analytically verified in literature [9]. Requirementtwo, statistical independency, is probably never completely true.However, in practice this dependence tends to be negligible, asdemonstrated in [11]. Section II will introduce and validate theparameterization scheme, associated with the third requirement,in more detail.

In Section III, a Kalman filter is tailored to estimation of .The Kalman filter is designed to estimate changes in a systemin which some prior knowledge on noise and system structure isavailable. The filter minimizes the mean squared error of the pa-rameters in a given system. Since the structure of the system inthis study is known, as explained in Section II, such a Kalmanfilter is used. The trade-off between speed of filter adaptationand estimation variance is explained. It is described how thebasic Kalman filter can be tuned to the requirements of eyemovement correction.

In Section IV, simulated data are used, in which andare known and combined in order to simulate . By com-

paring these to and it will be shown that the newmethod separates the components accurately. These compar-isons are made for various types of eye movement to illustratethat the method still performs very well, even in EEG signals


Fig. 2. Potentials e(p ; p ), e(p jp ), and e(p jp ) for experimental and simulated data. In each plot of e(p ; p ), the values for p and p where the crosssections are made are indicated.

with large eye movement artifacts. An important question tobe answered is how these results compare to results of existingtechniques. In this paper, results will be compared objectively tothe regression, SOBI and PCA methods. As a yardstick, a signalto noise ratio, signal-to-noise ratio (SNR), is defined.

Next in Section V, the eye-tracker method is applied to ex-perimental data. The estimated potential is displayed fordifferent types of eye movement. As there is no reference on theexact OA in this case, a qualitative discussion on these results isprovided, in lieu of an objective evaluation.

II. SECOND-ORDER STRUCTUREThe correction method is based on the three key requirements

listed in the introduction. The third assumption, on the rela-tion between and , will be investigated here in moredetail.

An experiment is performed in which is changed in aregulated way. A participant is seated in front of a monitor,at a distance of 65 cm, the eyes horizontally aligned with thecenter of the monitor screen. The participant is asked to track amoving dot with his eyes, and instructed not to move his head.The dot appears for 0.5 s on 225 (15 15) different positionson the monitor, starting at the upper left position, going step-wise towards the upper right, then dropping one line and going

stepwise back to the left side. This pattern continues until thebottom right corner is reached. EEG is recorded at 512 Hz. Elec-trodes are placed according to the 10/20 system [12]. A similartest is performed in a simulated environment, using a model ofthe head as described in [13]. In this simulated environment,there are no cerebral potentials and the exact relation between

and can be computed. For the experimental data,all samples recorded while the dot was at one position are aver-aged for each dot position.

Results for both the experimental and the simulated situationare shown in Fig. 2. Because OAs are most profound frontally,the displayed results belong to Fz and Fp1 electrode positions.One of these, Fz, is on the midline, showing symmetry betweenleft-right ocular orientations. For each electrode position, threeplots are shown. The left one illustrates , and the tworight ones show cross sections of this plot where either the hori-zontal or the vertical pupil position is fixed, resulting inand .

Note in Fig. 2 that is arbitrarily scaled. Also note thatand range from to 0.22. This actually represents

the position of a dot on the monitor screen, withat the center of the screen. To convert this position on the screento would require a minor extra transformation. However,proving that the relation between orientation and potential can


be estimated from the data is identical to proving that this holdsfor the position of the dot. Furthermore, for small angles, therelation between position on the screen and can be assumedto be linear [14].

Clearly visible in Fig. 2 is the well organized structure of. Vertical movements have a greater impact on the

value of than horizontal movements for these electrodepositions. Most cross sections show a first-order dependence.Obviously the plots for simulated data show less variance thanplots for experimental data because the simulated ones do notcontain any cerebral component. In the experimental data,

has greater variance than . This is due to thetime between consecutive samples. The way the experiment isset up, is recorded within , whereasis recorded in . As the recording takes more time, itis more difficult not to move the head and, thus, the variance ofthe recording will increase. Another important explanation forthis variance is baseline fluctuation during the recording, e.g.,due to electrode drifts. These fluctuations are typically of lowfrequency and are hard to distinguish from changes in ,which occur once every 7.5 s in this experiment. In record-ings with high-amplitude drifts, can be completelyobscured by these drifts, and even can be affected.The results shown in Fig. 2 are based on data in which baselinefluctuations were visually not detectable.

The only cross section showing a second-order dependence isfor the Fz position. As this position is on the midline,

the distances to both the eyes are equal. Therefore, the effectsof negative or positive deflection of are identical, causingan optimum at .

For other electrode position, not shown in Fig. 2,and also appear to be (at most) quadratic functions.We infer that may be approximated by a second-orderfunction. The structure of such a function is given in by

(1)or, more compactly, by

(2)with and

.

The six parameters in (1) togetherdetermine the MODEL CONDUCTION block in Fig. 1.

In (1), it can be seen that the impact of does not de-pend on , making it strange that this parameter is involvedin estimating an OA. Parameter , however, is importantin removing baseline fluctuations from . These fluctuationsmay change and are independent of . By including thisparameter in vector , can be estimated whilecorrecting for baseline fluctuations. The impact of can becompared to that of a high pass filter. Section III-C. will explainhow is tuned to track these fluctuations.

III. IMPLEMENTING A KALMAN FILTERA Kalman filter is capable of estimating the state of a system,

even when the exact structure of the modeled system is un-

known. The relation between and may be consid-ered as such a system because the structure for this relation isknown from Section II while the influence of physical parame-ters on the system, as given in vector , is unknown.

A. Basics of the Kalman FilterA Kalman filter estimates in a system that can be de-

scribed by

(3)with matrix representing the expected changes in overtime and representing fluctuations in that are white,with normal distribution, and independent of . The notation

is used to indicate the value of at the previoussample.

In this paper, represents the state of all unknown param-eters that influence the relation between and . As inmost applications of the Kalman filter, cannot be measuredand, therefore, can only be obtained indirectly by recording ,as seen in

(4)Because is unknown, it can only be approximated by anestimate . The Kalman filter in effect is a set of equationsthat minimize, defined as the a posteriori estimate error covari-ance

(5)

By minimizing a statistically optimal , with respectto any quadratic function of estimation error, is obtained. Withthis estimate , the ocular component can be removed from

completely. Since the system in this study is influencedby a noisy signal , it takes an infinite number of iterationsbefore the estimated is completely stabile. Given only alimited number of data, the Kalman filter always returns the bestpossible estimate for .

In a system in which the exact system structure is known,in which all noise characteristics are known, and in which aninfinitely long adaptation period is allowed, the Kalman filterwill achieve perfect separation between artifact and signal. Inpractice, the allowed adaptation time is limited and the noisecharacteristics are often not known exactly known. How thisinfluences the filter properties is explained in the next section.

The full Kalman filter equations are listed in (6)(10)(6)(7)(8)(9)

(10)with being the Kalman gain, the a priori estimateerror covariance, the identity matrix, the expected covari-ances of , and the expected variance of . More detailsand a derivation of these equations are given in [15].


B. Adaptation Time Versus Estimation Accuracy

The matrix and scalar are not defined by the Kalmanequations, but by the dynamic properties of the system that isbeing modeled. They may be set based on prior knowledge onthe dynamics or based on specific desires for the filter perfor-mance as will be explained in this section. In a Kalman filter,the time required for stabilization of is influenced by boththe choice of and the choice of . If is large comparedto , then (3) shows that will be determined mostly by

. As contains no information on , the adaptationof is slow. If is small compared to , then (3) showsthat will be determined mostly by . In thiscase, it seems fair to adapt fast because new informationon is available. Thus, from (8), small leads to a large

and, thus, faster adaptation.After sufficient iterations, will stabilize. The Kalman

gain will still depend on and, therefore, the fluctuationsof after stabilization also depend on . These fluctuationsinfluence the variance in . A similar reasoning applies for

. The choices of and are always the result of a trade-offbetween adaptation time and estimation accuracy.

It is important that is small compared to thesignal of interest, . Because represents thepart of that is not of cerebral origin, it will be referred to asnoise This noise should not be confused with what is known aselectrode noise, an extra additive component to signal thatis caused by the recording equipment. What is, depends onthe specific EEG study that is being performed. Clearly, whenthe signal of interest is , more noise can be tolerated thanwhen the signal of interest is only a small component ofas is the case in, e.g., evoked potential, EP, studies. By using asignal to noise ratio defined as

(11)

the performance of the correction can be objectively deter-mined. In (11), represents the start of the period overwhich SNR is calculated and represents the duration of thisperiod. For ongoing EEG studies, will be the start of themeasurement and the length of the measurement. For EPstudies, SNR can be determined for each separate stimulus,with the stimulus time and the (expected) duration ofthe stimulus response. For (11), it is required that and

are known. Therefore, SNR can only be determined forsimulated data. As depends on the choices for and

, it is clear that SNR also depends on these choices.An optimal correction can be achieved after an infinitely long

adaptation, assuming the vector stays constant and that thespectrum of is white. However, in practice, the level of cor-rection is restricted by the duration of the measurement, the fre-quency content of the EEG, and by the requirement that smallchanges in should be tracked. The physical processes thatchange are slowly varying and can be considered constantfor several minutes, e.g., temperature and sweating. Changesin vector can also be due to, e.g., electrode movementor small head movements. These will cause faster and greater

changes in . By instructing the participant not to moveduring the recording, these changes can be decreased. It will,however, still be necessary to track the remaining changes likebaseline fluctuations. Typically these only contain very low fre-quencies, below 0.1 Hz. In the recordings that are used in thispaper, baseline fluctuation is detected visually. The maximumfluctuation frequency is app. 0.03 Hz. Therefore, in this study, itis required that changes in can be tracked within 30 s. Thisvalue is fast enough to track changes in . A calibration pe-riod of 30 s should precede each experiment. Note that changesdue to fast head movements cannot be tracked in this manner.

C. Tuning the Kalman FilterIn (3), reflects how may change due, e.g., to tem-

perature changes of the subject, or electrode movements anddrifts. Since there is no prior knowledge on what changes toexpect, is set to the identity matrix . Alternatively, setting

equal to can be used to keep from unboundedgrowth, although no extreme values were seen in this study with

. The expected variance of is given in . The basicKalman filter assumes to be white whereas in reality,is known to be colored as the frequency content of the EEGis generally below 50 Hz. For now we will assume to bewhite and use the basic Kalman filter. As the average varianceof an EEG recorded with eyes open is approximately 144 ,as determined experimentally, is set to this value. It is knownthat choosing is generally more difficult. As we typically donot have the ability to observe directly, it is also difficultto estimate the variances of these parameters. Because of thisdifficulty, the structure of is simplified by assuming that thefluctuations in the parameters in are not correlated witheach other. This implies that all off-diagonal entries in arezero. The diagonal entries of indicate the variances of the sixparameters in . They are set to track changes in a limitedamount of time. As changes that influence conductance are ex-pected to affect all parameters simultaneously, all diagonal en-tries should be set to obtain a similar tracking time of 30 s forthe corresponding parameter. If the data is not rich enough it isnot possible to estimate and, thus, track some of the parameters.If for instance there are no eye movements, only can beestimated. For the Kalman filter, the tracking time is determinedby , and . The first two are fixed and, thus trackingtime is set to 30 s by setting . In Fig. 3, it is illustrated how thetracking time for depends on the fourth diagonal elementof , denoted as .

Because larger values for result in larger estimationvariance, is set to the smallest possible value that sta-bilizes within 30 s; . This is repeatedfor all six diaginal entries of , resulting in

. Stabilization of istypically determined by the slowest tracking parameter in

. Setting for each element separately will lead to equalstabilization within 30 s and will result in minimal estimationvariance for a given tracking speed. The risk of adaptive pa-rameter estimation is that brain activity that correlates in timewith eye movements is being removed as well. Although it isimpossible to eliminate this risk completely, we note that such


Fig. 3. Influence of on tracking time. Each subplot is initialised at 100(V=m). Results are obtained using simulated data with random eye move-ments.

brain activities are usually reflected in brief evoked potentials,whereas the change in ocular orientation as caused by eyemovements lasts much longer. The parameter tracking, withan adaptation speed of 30 s will hardly be modified by thecorrelation between brain activity and the change in . In[11], the influence of ocular orientation on brain activity isfound to be negligible.

IV. SIMULATED DATA, ACCURACY OF THE NEW METHODCOMPARED TO EXISTING METHODS

The most important feature for any OA removal method is itsperformance on experimental data. Objective classification ofremoval performance based on this type of data is, however, verydifficult or even impossible. A way to estimate this performanceis to use simulated data. In order to obtain meaningful results, itis important that the simulation model is a good representationof reality.

A. Simulated DataWe simulate EEG data using a boundary-element-method

(BEM)- based model of the human head. Brain activity is sim-ulated in this model by rotating a fixed number of brain-dipoleswith fixed dipole strength, and fixed position. This results in

with most of its signal power below 30 Hz, and havingan exponential decay of power with increasing frequency. Theexact properties of these dipoles, the model of the head, andproperties of the dipoles used to simulate eye movements andEEG are described in detail in [13]. The main advantages of thismodel lie in the ability to simulate separately, but realistically,the sources that are generating and using dipolemodeling. A minor drawback is the inaccuracy in the modelingof the skull. Whereas a skull has holes behind the eyes, themodel only simulates a closed-surface skull. This might leadto a slightly inaccurate representation of scalp topography ofthe OA. However, properties of topography are not used in themethod presented here, and they do not influence OA removal.EEG is simulated at all positions of the 1020 system. Be-cause the properties that determine conduction remain constantduring the whole simulation, is constant.

Next to the brain activity and eye movement artifact, elec-trode noise is also simulated, as described in [13]. Each elec-

TABLE ISIMULATED EYE MOVEMENTS TYPES AND THEIR DESCRIPTIONS

trode has some intrinsic noise which we simulated to be nor-mally distributed, with zero mean and a standard deviation of1 . This is similar to electrode noise levels as determinedfor the active electrodes we used in our experiments. Since thiselectrode noise is small compared to the simulated brain activity,small changes in standard deviation are not expected to result insignificant changes in correction performance.

Additional to EEG data, eye-tracker data also needs tobe simulated. Clearly will depend on the eye movement thatis simulated. To simulate eye movements, the ocular dipoles,implemented in the BEM model, are oriented towards a dot thatmoves over a screen positioned 0.8 meters in front of the sim-ulated head. As the dot moves, the ocular dipoles rotate, andan eye movement is simulated. The center of the screen is hori-zontally aligned with the eyes and is 45.45 . An eye trackerpositioned near the screen and directed to the head records .The vector will resemble the data on the movement of thedot. Therefore, the moving dot data are converted to simulate

. For the conversion, the 2-D position in centimeters is con-verted linearly to .

For this paper, four different types of data sets are simulated.Each of these sets can be characterized by a specific sort of eyemovement. This is done because any information on needsto be derived from the data. If the data for instance only containsone steady ocular orientation, it is not possible to accuratelyestimate all parameters in . By using different sets, we canevaluate how well the OA performs for these situations.

The types of eye movement that are simulated in the differentdata sets are listed in Table I. These types are chosen becausethey represent relevant eye movements that can occur duringexperiments. For random eye movements, the eyes are on av-erage directed towards the center of the screen. Because the fre-quency of eye movements is physically limited, e.g., it is onlypossible to focus on approximately three different spots withinone second, positions and are simulated by applyinga low pass filter, with a cut off frequency at 3 Hz, to a random,white noise signal. The second type is less realistic, but illus-trates what happens if there is no eye movement at all. In thiscase, an OA correction method should only remove a dc offset.The third type simulates a dot-tracking task where the dot makesa smooth circular movement over the screen. The fourth typesimulates a dot-tracking task where the dot jumps once every 2s from one point of the screen to another.

In Table I, is the maximum amplitude of vector withrespect to the center of the monitor screen ( ),represents the angular frequency of the deterministic eye move-


ment in rad/s, and indicates a normal-distribu-tion around zero with a standard deviation of . Note that thesamples did not exceed in our simulations. Samples for typethe saccade movement are drawn from , with equalprobability for the three entries.

The amplitudes of and can be scaled in the sim-ulation by increasing the strength of the sources. This way bothpotentials are scaled to realistic values. The amplitude ofis scaled to an experimentally determined value for each specificelectrode position. Note that the value for , which was set inSection III-C, was already based on these realistic values.

For each type, 40 s of data are simulated at a rate of 256 Hz.The first 30 s are used for parameter initialization. The last 10 sare used to evaluate the correction method.

The signal of interest in this case is . The SNR provides agood indication of how well the correction method succeeds inremoving the OA. In the following sections, results are shownfor the Fp1 position. This position is very close to the eyes and,therefore, contains the largest OA. Thus, results for this positioncan be considered as worst case results.

B. Comparison With Other MethodsIt is important to see how these results relate to those of ex-

isting techniques because this can verify whether the extra ef-fort of including an eye tracker in an experimental setup doesimprove OA removal. In this paper, we consider three existingOA removal methods.

Regression [16]:Bipolar recordings of electro-oculographic (EOG) arescaled and subtracted from the EEG. The number of EOGchannels that results in the best correction is still underdebate. In this paper, the best result of using either two,four, six, or eight EOG electrodes and, hence, one, two,three, or four bipolar EOG recordings, are given. BipolarEOG recordings have the advantage that brain activity thatis recorded by both electrodes is greatly reduced in theEOG and the bipolar EOG is, thus, relatively clean whencompared to a common reference EOG. However, brainactivity originating close to the eyes will still be visible ina bipolar EOG.Often a calibration period precedes the data that should becorrected. The regression parameters are calculated overthis period and subsequently fixed and used to correct thedata of interest. Here, in all but one of the simulated eyemovement types, the eye movements cover a wide range,similar to a range that would have been used during calibra-tion. Adding a calibration period would, therefore, not en-hance performance. The results shown here may be viewedas correcting calibration data with coefficients found on thesame calibration data.Because many eye movements are made during each type,the coefficients will be accurate. Only the type 1 movementis likely to improve significantly when a calibration trial isadded.

PCA: Principal Component Analysis [17]:All recorded EEG, and EOG, channels are projected ontoa new set of othogonal base vectors in an attempt to decor-relate brain and ocular activity. The vectors that resemble

the EOG channels are removed and the remaining vectorsare used to construct a cleaned EEG.

SOBI: Second Order Blind Identification [18]:Component analysis technique similar to PCA, but with theability to exploit autocorrelation in brain activity and auto-correlation in ocular activity. In [19] and [20], it is shownthat SOBI can be tuned very accurately for the purpose ofextracting small components arising from the primary so-matosensory cortex. Next to this, SOBI is capable of re-moving OAs and other artifacts, making it an easy to useand versatile algorithm. For the SOBI method, a numberof time lags need to be chosen. The components retrievedby SOBI are uncorrelated with each other. The correla-tions between all components are calculated at these spe-cific time lags. The SOBI algorithm minimizes the sum ofall calculated correlations, excluding all autocorrelations.Unfortunately the study as to which lags should be usedto optimize SOBI performance is currently limited to ex-tracting primary somatosensory cortex signals [20]. Theoptimal setting of lags depends on spectral properties of thesignals that need to be separated and is, therefore, likelyto be different here. In this paper, we use lags of 1, 2, 3,5, 10, and 20 samples at 256 Hz. These lags are equal tothose used in [13], where SOBI was found to be the bestperforming correction algorithm. The SOBI method may,however, improve if the optimization of the selected lagsfor OAs is studied in more detail.

These three methods are selected because, in previous studies,they have been shown to perform well in OA removal [10], [13].

In order to make a fair comparison between the methods,it is important that the effective amount of data available toall methods is equal. The eye-tracker method requires a 30-sparameter tracking period prior to the 10 s we use for eval-uation. Even though this is merely a parameter initializationcomparable to calibrating regression-based approaches, this pe-riod provides the eye-tracker method with extra informationthat should also be available to the other methods. In order tomake a fair comparison, the other methods should, thus, esti-mate based on a similar amount of data. The eye-trackermethod uses an adaptive filter. Therefore, the data of these 30 sis weighted exponentially.

Because the other algorithms use datasets of fixed lengthand apply the same weight to each sample, the length of thesedatasets, , should be matched to the amount of data in theexponentially weighted 30 s. This relation is given by

(12)

Stabilization of occurs within three times the decay con-stant, , of this equation. Since this corresponds to 30 s,

, and from (12) .For a 40-s data-segment, this means that in order to have

matching amounts of data, the eye-tracker method starts cali-brating at the start of the segment whereas the other methodsshould operate on the last 20.5 s of the segment. Finally, the last10 s are used to evaluate the correction methods.

As discussed in the introduction, most correction methods as-sume a stationary relation between eye movement and recorded


TABLE IIPERFORMANCES OF DIFFERENT OA REMOVAL METHODS FOR DIFFERENT TYPES OF SIMULATED EYE MOVEMENTS. PERFORMANCE CALCULATED OVER 10

SECONDS IS EXPRESSED AS SNR. RESULTS ARE AVERAGED OVER 18 SIMULATED SETS

artifact as well as a stationary relation between cerebral activityand recorded EEG. In order to comply with this assumption,epochs of short duration are often used. Increasing the epochduration for such methods, to compensate for the parametertracking period, might actually be a disadvantage. Because inthe simulated data all parameters in are constant the assump-tion of stationarity is not violated, so there is no need to useshorter epoch duration and (additional) difficulties when re-at-taching different epochs are avoided.

C. ResultsAll four data-types were corrected by the eye-tracker-based

method and also by the three other OA removal methods, theSNR calculated for all methods are shown in Table II. The re-sults are averaged over 20 simulations to decrease the influenceof the randomness of on the results.

By performing a two way analysis of variance (ANOVA) onthese results, their significance is tested. The p-value for themethod effect is very small ( ). This is a strong in-dication that the mean SNR value is different for the correc-tion methods. The p-value for the type-effect is 0.0053, which isalso highly significant. This indicates that for some types of eyemovement results are better than for other types. There is, how-ever, also a strong interaction between eye movement type andcorrection method effect, making the results for the type-effectsomewhat questionable. By using an ANOVA, we make threeassumptions on the data. First we assume that the SNR is nor-mally distributed for each correction method, which is approxi-mately true. Second, the standard deviations of the SNR for thedifferent methods are assumed to be identical. Although this isnot true, the consequences are small because each method hasa similar number of SNR values. Finally it is assumed that thesimulations are independent.

D. DiscussionThe eye-tracker method consistently outperformes the three

other methods, often by a large margin. In any of the four typesof eye movement, the eye-tracker method results in an averageSNR of over 10 dB. This indicates that power of the remainingartifact after correction is approximately 10 times smaller thanthe power of the estimated . The best SNR is obtained fortype 2. For this type, the optimal correction would be to re-move only dc because the eyes did not move during simula-tion. However, prior to applying any correction method, the dc

is already removed by subtracting the mean of the signal be-cause this is common procedure with most EEG recordings.Therefore, type 2 illustrates how the correction methods affectSNR if no correction is needed. Without correction the SNRin this case is infinitely large since there is no noise. All cor-rection methods slightly distort because all SNR values inTable II are finite. The other three methods result in a lowerSNR than the proposed eye-tracker method. For the regressionmethod, this is probably caused by the presence of a small cere-bral component in the recorded EOG, due to volume conduc-tion. For all eye movement types, the optimal regression wasachieved when two bipolar EOG recordings, and hence fourEOG electrodes, were used. The PCA and SOBI identify sev-eral different components the data of type 2 movement, but theylack the clean reference signal of an eye tracker. Therefore,the EOG channels are used to determine which of the com-ponents are ocular and these channels will not show the eyemovement as clear as . For types 3 and 4, the SNR of theSOBI method is close to the SNR of the eye-tracker method.The SOBI method exploits deterministic time structures thatare hidden in the data, and it is, therefore, not surprising thatunder these circumstances SOBI achieves an SNR of 15.6 and9.9 dB, respectively. The use of in the eye-tracker methoddoes, however, perform slightly better. Indicating that the extradata that is available due to the use of the eye tracker, providesimportant extra knowledge on the OA. The PCA and the re-gression method are consistently outperformed by SOBI andby the eye-tracker method. For PCA, this is probably becausePCA does not use the temporal information of , whereasSOBI does [18]. Regression has the risk of overcorrecting andremoving brain activity from the EEG. As discussed, EEG reg-istrations often contain baseline fluctuations due to, e.g., elec-trode drifts. The eye-tracker method is tuned to track these fluc-tuations. Even though in this simulated data there are no fluctu-ations in parameters, the performance of the eye-tracker methodis still better than the other methods. If the data would havecontained baseline fluctuations, the eye-tracker method wouldhave removed them with essentially no impact on SNR. Becausethese fluctuations are mostly not correlated for different elec-trodes, both SOBI and PCA would not remove them and, there-fore, the SNR shown for these methods in Table II would de-crease. The results in Table II are based on 20 simulated datasets.The deviation of single experiment results around this averageare small and, thus, indicate reliable correction. For the type 1


movement, e.g., only one simulation deviated by 1.8 dB (havingan SNR of 13.7 dB), all other simulations deviated by less than1.1 dB.

For the eye-tracker method, a 1-channel EEG recording issufficient to obtain the results that are shown in this study. Theextra requirement for this method is that an eye tracker is addedto the experimental setup. For the SOBI and PCA method, noeye tracker is required, but EOG electrodes need to be includedin the measurement as well as a larger number (21) of EEGchannels to obtain the results presented here.

V. EXPERIMENTAL DATA, USING THE NEW METHODFOR CORRECTION

EEG, EOG, and eye-tracker measurements are collected from9 participants aged 1921, 5 male and 4 female. The partici-pants performed a task involving eye movements. During thistask, a (moving) dot appears on a 19-in monitor. Like with thesimulated data, there are different types of eye movements, cor-responding to types 24 from Table I. The participant is askedto keep his eyes on the moving dot. An eye tracking system ispositioned directly below the monitor and records the positionof the participants left eye. It uses an infrared light and from thelight reflected by the eye, the position of the center of the pupilis determined. EEG measurements are performed with 21 EEGelectrodes positioned according to the 1020 system. Another6 electrodes are used to record the EOG. They are positionedabove and below both eyes, left of the left eye and right of theright eye. Recordings for all electrodes are offline referenced tolinked mastoids. EEG and EOG are recorded at 256 Hz usingthe BioSemi ActiveTwo system with sintered Ag/AgCl elec-trodes using a lowpass filter with a cutoff of 67 Hz. Eye-trackerdata are recorded at 50 Hz using the SensoMotoric InstrumentsRED eye tracker with an angle resolution better than 0.1 . Theeye-tracker data are up-sampled from 50 to 256 Hz afterwardsand synchronized with the EEG recording. During the task, theparticipant sits comfortably in front of a monitor at 0.8 m dis-tance with the head supported and eyes are horizontally alignedwith the center of the screen.

A. ResultsWith experimental data it is not possible to calculate the SNR

because is unknown. For this reason, the estimated EEG,, is presented as a result of correction. The segment that is

shown corresponds to the 10 s immediately after the 30-s initial-ization period. Because SNR cannot be determined, it is not pos-sible to objectively prove the accuracy of these corrections. Re-sults, again shown for the Fp1 position, are illustrated in Fig. 4for the eye tracker and the SOBI method as these performed beston simulated data. The matrix is set to make sure that all pa-rameters in stabilize within 30 s.

B. DiscussionWhen is observed for these three situations using the

eye-tracker method, the only remarkable disturbance that is stillclearly visible in is in the saccade data. Here there remainsome small disturbances directly after the onset of each saccade.The disturbance is also visible after SOBI correction and, there-fore, it is unlikely that this disturbance is caused by the correction

method itself. The disturbances are probably caused by what isknown as the rider-artifact [21]. This artifact is often seen withsaccidic eye movements and is caused by slight changes in thepositions of the eye lid that occur when saccadic eye movementsare made. Because the eye tracker did not monitor eye lid move-ment, this type of artifact could not be removed by the eye-trackermethod. The artifact is mostly seen with vertical eye movements.When the SOBI results are compared to the eye-tracker results,it appears that the eye-tracker method is better in removing thebaseline drifts from the data. These drifts can be uncorrelatedbetween electrodes, and because SOBI relies on inter-electrodecorrelations to detect artifacts, the SOBI method will not correctthem as well as the eye-tracker method. It should be noted thatfor consistency with the simulated data, the regression methoduses only two bipolar EOG recordings. This corresponds to theoptimal setting found on simulated data. On experimental datait can be argued that a third recording is required to compensatefor blinks and small eyelid movements. However, adding athird EOG recording increases the risk of overcorrection andremoving brain-related activity from the EEG.

VI. DISCUSSION AND CONCLUSIONAs eye movement artifacts are often seen in EEG recordings,

correction for these artifacts is frequently needed to get a clearindication of the electrical activity of the brain. For the purposeof correction, it is desirable to have an indicator of which eyemovements were made during the recording. Often the EOG isused for this purpose, however, in this study we introduce the useof an eye tracker to monitor eye movements. This has the advan-tage that the data recorded by the eye tracker cannot be corruptedby any electrophysiological signals. Using the eye-tracker dataas a reference for correction, thus, is potentially very powerful.Nevertheless two issues related to the eye tracker should beaddressed. Firstly, the eye tracker used in the experiments, isnot able to distinguish between small head movements and eyemovements. It only represents the position of the pupil in a fixedframe. Although this implied for our experiments that the partic-ipants had to be specifically instructed not to move their head,in the future the use of a head mounted eye tracker can avoidthis constraint. During the experiments the position of the headwas continuously monitored and no large movements were de-tected. Secondly, the eye tracker is not capable of detecting pupilposition while the eyes are closed during blinks or during pe-riods of prolonged eye closure. For prolonged closure periods,like during sleep, this implies that the eye-tracker-based methodcan not be used. For blinks, it implies that the pupil position in-formation will be briefly interrupted causing a gap in the pupilinformation. Although in the current study such gaps were notpresent by selecting segments without blinks, in most data theywill occur. For such a brief period, it seems fair to halt adapta-tion, and continue once the pupil position can be recorded again.The properties that determine the parameters in are ex-pected not to change significantly in these brief periods.

By using both simulated and experimental data, it is deter-mined how the orientation of the eye, recorded by the eyetracker, influences the EEG signal, , that is recorded at anelectrode. This relation, described in (1), is at most of second-order.


Fig. 4. Experimental data for three different types of eye movement in the first row of each subfigure, recorded at Fp1. The estimated c^(t) corrected by theeye-tracker method and the SOBI method are in, respectively, the second and third row of each subfigure. Note the different scales of the y-axis, which are usedto show the full range of r(t) while also showing some detail in the corrected data.

Knowing the order of the relation between and , aKalman filter is used for obtaining the parameters that specifythe exact relation for each electrode position. The Kalman filteris an adaptive filter that can estimate these parameters and tracktheir changes in a limited period of time. In this paper, we usea tracking time of 30 s after which the estimated parametersshould be stabilized.

To gain insight in how well this new correction method per-forms, it is applied to both simulated and experimental data.The same data are also corrected by three established correc-tion methods, which have been previously reported to result inaccurate OA removal. Different types of eye movements are an-alyzed in this study because the morphology of is likely toinfluence the performance of some correction methods.

On simulated data, the eye-tracker method performs verywell. When the method is tuned optimally for this data, the SNR

after correction is over 10 dB for all types of eye movements.When compared to the other three methods, only the SOBImethod shows similar results for one eye movement type. Fordata containing baseline fluctuations, often seen in recordings,it is expected that the eye-tracker method will perform signif-icantly better than the SOBI method because the eye-trackermethod has the ability to remove this type of artifact.

It should be noted that low frequency components in alsocause slow changes in . These will affect and, there-fore, influence . This has a negative effect on the correc-tion. The estimate is correlated to , while isindependent of . After stabilization, the difference

is not white noise but may display considerablefluctuations. These fluctuations are proportional to fluctuationsin because of the small cerebral influence on . Theamplitude of these fluctuations depends on the first diagonal ele-


ment of matrix , denoted as . Larger will lead to higheramplitude fluctuations in . For the tracking time used in thispaper, the amplitude of these fluctuations is found to be negli-gible. When the new method is applied, e.g., to data containing alot of blinks, it is desirable to have a much shorter tracking timeafter each blink. If a shorter tracking time is required, the fluc-tuations in need to be considered. For now, a Kalman filterthat assumes that is of white spectrum already outperformsthe other methods, even though in simulations and reality thisspectrum is not white,. The assumption of matrix being diag-onal is, as mentioned in Section III-C a simplification of reality.This assumption is necessary because the true interactions be-tween parameters are very difficult, if not impossible, to assess.Therefore, a choice needs to be made on how to implement ma-trix . By using the simplest scenario, of diagonal , adequatecorrection is already achieved, although there is still room forimprovement.

On experimental data results for the different correctionmethods also look convincing. When inspected visually, the

that is estimated by the eye-tracker method appears to be aclean EEG signal that does not show obvious ocular influencesany more. The only exception to this is a small change inpotential that is seen for saccadic eye movements. This changeis probably caused by a rider-artifact that starts simultaneouslywith the start of some saccades. As this artifact is known to becaused by movement of the eyelid, it cannot be removed by thenew correction method. The results for the SOBI method appearto contain small baseline fluctuations that are not corrected for.

It should be noted here that components-based methods areoften praised for: 1) their ease of use; 2) their ability to removeeye movement artifacts as well as blink artifacts; 3) their abilityto extract small specific brain activities, like EPs, in trial-basedstudies accurately. In this paper, the two component-basedmethods are outperformed by the eye-tracker-based method.Nevertheless, our study does not distinct in any way on howwell any correction method will be successful in extractingsmall EP signals. In this paper, can be affected by proper-ties, like retinal luminance, that influence the relation betweenthe corneo-retinal dipole and recorded signal. Some of theseproperties will not affect the relation between a brain activitydipole and the recorded signal, and hence the fluctuations inthese relations parameters will be different. Deciding whichmethod is best at detecting EPs requires a different study, andcould be combined with an OA removal method.

For the eye-tracker method, a 1-channel EEG recording issufficient to obtain the results that are shown in this study. Theextra requirement for this method is that an eye tracker is addedto the experimental setup. For the SOBI method, no eye trackeris required, but EOG electrodes need to be included in the mea-surement as well as a larger number of EEG channels to obtainthe results presented here.

Currently, the applications for which this method can be usedare greatly limited by the inability to correct blink artifacts. Inorder to increase the variety of applications the new method canbe used for, a more advanced eye tracker which also monitorsmovement of the eyelid can be used. Such an eye tracker mayalso be used for the removal of blink artifacts. The use of ahead-mounted eye tracker can eliminate the strict need to avoid

any head movements during the recording. Some head-mountedeye trackers can be worn like glasses and do not interfere withelectrode caps. An extra problem that will arise concerns pupilposition detection during a blink and during periods of pro-longed eyelid closure. The SOBI method and the PCA methodare based on correlation between different electrode positions.Because these two methods do succeed in removing part of theOA, this correlation does contain information that is relevant fordetermining which part of is caused by eye movement. Theeye-tracker method is not yet able to use this extra informationbecause it is based on a 1-channel recording. The method can,however, be extended and improved to deal with multiple elec-trodes and the covariances between the different for theseelectrodes. In summary, if eye movement artifacts need to be ac-curately removed from EEG signals, especially for demandingapplications such as single-trial-based experiments, the use ofan eye tracker during experiments is essential.

REFERENCES[1] C. H. M. Brunia, J. Mocks, M. M. C. van den Berg-Lenssen, M. Coelho,

M. G. H. Coles, T. Elbert, T. Gasser, G. Gratton, E. C. Ifeachor, B. W.Jervis, W. Lutzenberger, L. Sroka, A. W. van Blokland Vogelesang, G.Van Driel, J. C. Woestenburg, P. Berg, W. C. McCallum, P. D. Tuan,P. V. Popock, and W. T. Roth, Correcting ocular artifacts in the EEG:A comparison of several methods, J. Psychophysiol., vol. 3, pp. 150,1989.

[2] R. J. Croft and R. J. Barry, Removal of ocular artifact from the EEG:A review, Neurophysiologie Clinique, vol. 30, no. 1, pp. 519, 2000.

[3] T. P. Jung, C. Humphries, T. W. Lee, S. Makeig, M. J. McKeown,V. Iragui, and T. J. Sejnowski, Removing electroencephalographicartifacts: Comparison between ICA and PCA, Neural Netw. SignalProcess., pp. 6372, 1998.

[4] R. Verleger, Should we really use different estimates for correctingEEG artefacts produced by blinks and by saccades?, J. Psychophysiol.,vol. 14, no. 4, pp. 204206, 2000.

[5] P. Berg and M. Scherg, Dipole models of eye movements and blinks,Electroencephalogr. Clin. Neurophysiol., vol. 79, pp. 3644, 1991.

[6] K. G. Rottach, V. E. Das, W. Wohlgemuth, A. Z. Zivotofsky, and R. J.Leigh, Properties of horizontal saccades accompanied by blinks, J.Neurophysiol., vol. 79, pp. 28952902, 1998.

[7] M. Iwasaki, C. Kellinghaus, A. V. Alexopoulos, R. C. Burgess, A. N.Kumar, Y. H. Han, H. O. Lders, and R. J. Leigh, Effects of eyelid clo-sure, blinks, and eye movements on the electroencephalogram, Clin.Neurophysiol., vol. 116, pp. 878885, 2005.

[8] F. Matsuo, J. F. Peters, and E. L. Reilly, Electrical phenomena asso-ciated with movements of the eyelid, Electroencephalogr. Clin. Neu-rophysiol., vol. 38, pp. 507511, 1975.

[9] M. S. Hamalainen, R. Hari, R. J. Ilmoniemi, J. Knuutila, and O. V.Lounasmaa, MagnetoencephalographyTheory, instrumentation,and applications to noninvasive studies of the working human brain,Rev. Modern Phys., vol. 65, no. 2, pp. 413497, 1993.

[10] G. L. Wallstrom, R. E. Kass, A. Miller, J. F. Cohn, and N. A. Fox, Au-tomatic correction of ocular artifacts in the EEG: A comparison of re-gression-based and component-based methods, Int. J. Psychophysiol.,vol. 53, no. 2, pp. 105119, 2004.

[11] C. W. Hesse, E. Seiss, R. M. Bracewell, and P. Praamstra, Absenceof gaze direction effects on EEG measures of sensorimotor function,Clin. Neurophysiol., vol. 115, no. 1, pp. 2938, 2004.

[12] H. Jasper, The ten twenty electrode system of the international feder-ation, Electroencephalogr. Clin. Neurophysiol., vol. 10, pp. 371375,1958.

[13] J. J. M. Kierkels, G. J. M. van Boxtel, and L. L. M. Vogten, A model-based objective evaluation of eye movement correction in EEG record-ings, IEEE Trans. Biomed. Eng., vol. 53, no. 2, pp. 246253, Feb.2006.

[14] J. Malmivuo and R. Plonsey, Bioelectromagnetism: Principles and Ap-plications of Bioelectric and Biomagnetic Fields. Oxford, U.K.: Ox-ford University Press, 1995, ISBN 0-19-505823-2.

[15] R. G. Brown and P. Y. C. Hwang, Introduction to Random Signals andApplied Kalman Filtering John Wiley, 1997, ISBN 0-471-12839-2.


[16] S. Chatterjee and A. S. Hadi, Influential observations, high leveragepoints, and outliers in linear regression, Statist. Sci., pp. 379416,1986.

[17] G. J. M. van Boxtel, Computational and statistical methods for an-alyzing event-related potential data, Behavior Res. Meth., Instrum.Comput., vol. 30, no. 1, pp. 87102, 1998.

[18] A. Belouchrani, K. Abed-Meraim, J. F. Cardoso, and E. Moulines, Ablind source separation technique using second-order statistics, IEEETrans. Signal Process., vol. 45, no. 2, pp. 434444, Feb. 1997.

[19] A. C. Tang, M. T. Sutherland, and C. J. McKinney, Validation ofSOBI components from high-density EEG, NeuroImage, vol. 25, pp.539553, 2004.

[20] M. T. Sutherland, J. Liu, and A. C. Tang, Temporal delays in blindidentification of primary somatosensory cortex, presented at the 3rdIEEE Int. Conf. Machine Learning and Cybernetics (ICMLC 2004),Shanghai, China, 2004.

[21] B. W. Jervis, E. C. Ifeachor, and E. M. Allen, The removal of ocularartefacts from the electroencephalogram: A review, Med. Biol. Eng.Comput., pp. 212, 1988.

Joep J. M. Kierkels was born in Baexem, theNetherlands, in 1979. He received the M.Sc. degreein biomedical engineering from the Eindhoven Uni-versity of Technology, Eindhoven, the Netherlands,in 2002.

He is currently working towards the Ph.D. degreeat the department of electrical engineering, Eind-hoven University of Technology.

His research interests include electro-encephalog-raphy, signal separation, model validation and artifactreduction.

Jamal Riani was born in Tetouan, Morocco, in 1977.He received his engineering degree from the EcolePolytechnique, Palaiseau, France, in 1999 and theEcole Nationale Superieure des Telecommunications(ENST), Paris, France, in 2001.

He is currently working towards the Ph.D. degreeat the department of electrical engineering, Eind-hoven University of Technology.

His research interests include signal processing fordigital transmission and recording systems.

Jan W. M. Bergmans (SM91) received the degreeof Elektrotechnisch

Ingenieur, cum laude, in 1982 and the Ph.D.degree in 1987, both from Eindhoven Universityof Technology (TU/e). From 1982 to 1999 he waswith Philips Research Laboratories, Eindhoven, TheNetherlands, working on signal-processing tech-niques and IC-architectures for digital transmissionand recording systems.

In 1988 and 1989, he was exchange Researcher atHitachi Central Research Labs, Tokyo, Japan. Since

1999 he is full Professor and Chairman of the signal Processing Systems Groupat TU/e. He has published extensively in refereed journals, has authored a bookDigital Baseband Transmission and Recording (Kluwer Academic, 1996), andholds some 30 U.S. patents.

Geert J. M. van Boxtel was born in Tilburg, TheNetherlands, in July 1958. He received the M.Sc. de-gree in psychology and the Ph.D. degree in 1989 fromTilburg University, in 1989 and 1994, respectively.

He was a Postdoc at the University of Amsterdam,Amsterdam, The Netherlands, and at the Universityof Dortmund, Dortmund, Germany. He is currently anAssistant Professor of Psychology at Tilburg Univer-sity, Tilburg, The Netherlands. His current researchinterests include cognitive neuroscience, especiallyevent-related brain potentials and its applications.

Dr. Van Boxtel has been a member of the Society for PsychophysiologicalResearch and the Federation of European Psychophysiological Societies since1989.

Download - Using Eye Tracker for Accurate Eye Movement Artifact Correction

Top Related