Automatic Removal of Ocular Artifacts in the EEG

without an EOG Reference Channel

German Gomez-Herrero1†, Wim De Clercq2∗, Haroon Anwar1, Olga Kara1, Karen Egiazarian1,Sabine Van Huffel2∗, and Wim Van Paesschen2�

1Tampere University of TechnologyInstitute of Signal ProcessingP.O. Box 553, 33101 Tampere

FINLAND†E-mail: german.gomezherrero@tut.fi

2Katholieke Universiteit Leuven∗Dept. of Electrical Engineering (ESAT/SCD) and

�Dept. of Neurology, University Hospital GasthuisbergBELGIUM

ABSTRACT

This paper presents a method for removing electroocu-lar (EOG) artifacts in the electroencephalogram (EEG).The procedure is based on blind source separation (BSS)and, in contrast to methods already available in the lit-erature, it is completely automated and does not requirethe availability of peri-ocular EOG electrodes. The pro-posed approach removed most EOG artifacts in 6 long-term EEG recordings containing epilectic seizures with-out distorting the recorded ictal activity.

1. INTRODUCTION

Eye movements and blinks produce electrical potentialsthat propagate over the scalp creating significant elec-trooculographic (EOG) artifacts in the recorded elec-troencephalogram (EEG).These artifacts often compli-cate the interpretation of the EEG. EOG artifacts ob-scuring the EEG at the time of seizure onset may beproblematic in the setting of a preoperative evaluation ofpatients with refractory epilepsy, since the ictal record-ings are crucial for the localization of the epileptogeniczone. Furthermore, EOG artifacts are a source of falsepositives in automatic seizure detection systems. Someepileptic seizures appear very infrequently and very longEEG recordings are necessary to capture them. There-fore, manual analysis of such huge amount of data is verytime consuming and fully-automated processing methodsare highly desirable.

Traditionally, EOG artifacts have been corrected us-ing regression-based methods which assume that signalsrecorded from electrodes placed near the eye collect cleanEOG free of EEG contamination. The corrected EEG isobtained by regressing out these reference EOG signalsfrom the signals recorded at the scalp electrodes [1–3].However, in an Epilepsy Monitoring Unit (EMU) theEOG is not always recorded simultaneously with the

EEG since the EOG electrodes are cumbersome for thepatient who is monitored for typically one week. Asa consequence, processing methods should not rely onEOG recording.

Other group of methods are based on decomposingthe recorded EEG signals into spatial components thatisolate the artifacts and the cerebral activity. Classicalapproaches for performing such spatial decompositionsare Principal Component Analysis (PCA) [4] and Sin-gular Value Decomposition (SVD) [5]. More recently,Blind Source Separation (BSS) techniques and speciallyIndependent Component Analysis (ICA) have becomevery popular for separating the EEG and EOG com-ponents [6, 7]. Apart from an accurate spatial decom-position, fully-automated component-based methods re-quire of robust criteria for identifying EOG-related com-ponents. Several criteria have been proposed in the lit-erature that require the availability of one or more EOGchannels (e.g. [2, 8]). In [9], a criteria based on Singu-lar Value Fraction (SVF) that did not require an EOGreference was used to select interesting components forepileptic seizure detection. However, the latter approachrequires the user to manually select an SVF thresholdwhich critically determines the accuracy of the selection.Furthermore, the authors of [9] recognize the difficultiesof finding a robust threshold. In contrast, our method istruly automatic since the user is not required to selectany critical analysis parameter.

2. METHODS

2.1. The EEG inverse problem

Typically, EEG observations are obtained at the outputof an array of scalp electrodes, where each sensor receivesa different combination of source signals of ocular (the”EOG” sources) and neural origin (the ”EEG” sources).

1-4244-0413-4/06/$20.00 ©2006 IEEE 130 NORSIG 2006

Let s(t) = [s1(t), ..., sn(t)]T be the n original unob-served sources, and let x(t) = [x1(t), ..., xm(t)]T be them mixtures observed at the electrodes. Here, as in thefollowing, T denotes transposition and t denotes the sam-ple index t = 1, ..., L. Let us call ΓEEG the set of indexessuch that si(t) ∀i ∈ ΓEEG are the source potentials ofneural origin. Similarly, let ΓEOG be the set of indexessuch that si(t) ∀i ∈ ΓEOG are the sources of ocular ori-gin. Most EOG correction methods assume that the sig-nal recorded at the jth electrode can be modeled as thefollowing instantaneous mixture:

xj(t) = xj,EEG(t) + xj,EOG(t)=

∑i∈ΓEEG

ajisi(t)++

∑i∈ΓEOG

ajisi(t)(1)

where aji is the transfer coefficient from the ith sourceto the jth scalp electrode. This model can be compactlyexpressed using matrix notation as:

x(t) = xEEG(t) + xEOG(t) = A(t)s(t)= AEEG(t)sEEG(t) + AEOG(t)sEOG(t) (2)

Using Eq. (2) we can formally express the EEG inverseproblem as the problem of estimating xEEG(t) from x(t).

2.2. EOG correction through spatial filtering

Spatial filtering methods address the EOG inverse prob-lem in three steps:

1. Estimate the mixing matrix A using a finite set ofobserved data x(t), t = 1, ..., L.

2. Identify the columns of A corresponding to artifac-tual and neural EEG components, i.e. identify thesub-matrices AEEG and AEOG from Eq. (2).

3. Recover the clean neural EEG activity by means ofthe following spatial filter:

xEEG(t) = AEEGA#EEGx(t) (3)

where # denotes the Moore-Penrose pseudoinverse.

Provided that the time courses of the ocular and neu-ral electrical activity are mutually independent, matrixA can be estimated using spatial ICA, which is a BSS ap-proach that considers the source signals as mutually inde-pendent random variables. An alternative BSS methodis Second Order Blind Identification (SOBI) [10], whichassumes that the source signals are temporally uncorre-lated to each other but they have non-zero time-delayedauto-correlations. Under these assumptions, SOBI com-putes the mixing matrix as the matrix that jointly di-agonalizes a set of p cross-correlation matrices R(τi) =E

[x(t)x(t − τi)T

], where i = 1, ..., p, and E[ ] is the

expectation operator. In this paper we use p = �L/3�where L is the number of EEG data samples.

In general, BSS makes very weak assumptions aboutthe mixing matrix, namely that it is of full column-rank.By contrary, PCA requires the columns of A to be mu-tually orthogonal which is very restrictive since the scalpactivity maps of EOG and frontal EEG sources are likelyto be non-orthogonal to each other. A major disadvan-tage of ICA-based methods is that they require the esti-mation of complex statistical measures of independencewhich makes them rather sensible to model inaccuracies.By contrary, SOBI is based on simple second-order sta-tistical quantities which are much easier to estimate andmore robust to modeling errors. Furthermore, severalneuroscientific studies support the use of SOBI for sep-arating EEG and EOG components [7, 8]. Because ofthis, the first step of our EOG correction approach con-sists on performing a SOBI decomposition of correlativeEEG frames (or data windows). By using this movingwindow analysis scheme we aim to partially cope withthe non-stationary nature of EEG signals. The lengthof the frames should be enough for allowing an accurateestimation of the mixing matrix coefficients but other-wise does not critically determine the accuracy of thecorrection. As a simple rule of thumb, we recommendusing as frame length at least 0.25 × m2 seconds, wherem is the number of scalp electrodes. In the correctionresults shown below we analyzed the data in correlativenon-overlapping frames of 200 seconds.

2.3. Identifying ocular components

The spectrum of ocular electrical activity is typicallycharacterized by few predominant low-frequency com-ponents. Such type of signals can be easily identifiedby a low fractal dimension (FD) which is a measure ofsignal complexity. By contrary, neural EEG traces aretypically characterized by a flatter and more spread spec-trum which accounts for higher FDs [11]. The FD canbe calculated using several methods. We decided to useSevcik’s algorithm [12] because it allows a fast computa-tion and is quite robust to the presence of noise.

If a waveform has coordinates (xi, yi), Sevcik’s algo-rithm first maps it into a unit square by x∗

i = xi

xmax,

y∗i = (yi−ymax)

(ymax−ymin) , where xmax is the maximum xi, andymax and ymin are the maximum and minimum yi. Then,the FD is computed as FD = 1 + ln(l)

ln(2(n−1)) , where l isthe total length of the waveform and n is the number ofpoints of the waveform.

In order to obtain a more robust estimate from non-stationary and noisy signals like the EEG we propose tocompute the FD of an EEG signal as the mean value ofthe FDs obtained in correlative frames of length equalto 10% of the total length of the signal. We will refer to

131

this estimated FD as mean FD (mFD). After analyzingvarious real EEG datasets we have confirmed that themFDs of EOG-like waveforms are consistently lower thanthose of EEG-like waveforms. However, the actual valuesof the mFDs depend on experimental factors such as thesampling rate and pre-processing filter used to acquirethe EEG data. Therefore, it is not possible to select auniversal mFD threshold below which a component is tobe considered of EOG nature. To overcome this problemwe propose the following approach for selecting the EOG-components among the components estimated by SOBIin each analyzed EEG frame:

1. Sort the components according to increasing val-ues of their corresponding mFD. Let us denote thesorted components by s(1)(t), ..., s(N)(t) and theircorresponding mFDs by φ(1), ..., φ(N). Note thatthe number of components estimated by SOBI isN = min(n,m) being m the number of data chan-nels and n the number of hidden sources. Note alsothat for A to be of full column-rank it is necessarythat m ≥ n.

2. Identify as EOG components s(1)(t), s(2)(t), ..., s(k)(t)where k is the smallest integer in the range �N/2� ≥k > 1 such that

(φ(k+1) − φ(k)

)<

(φ(k) − φ(k−1)

).

If there is not any k in the specified range satisfyingthe required condition then k = 1.

3. RESULTS

Six long-term EEG recordings, recorded at an EMU,were used in this study. The data was collected from21 scalp electrodes placed according to the international10-20 System with additional electrodes T1 and T2 onthe temporal region. The sampling frequency was 250 Hzand an average reference montage was used. The EEGrecordings contain ictal activity from patients with MesialTemporal Lobe Epilepsy (MTLE). Fig. 1 (a) shows a10 seconds EEG epoch of one of the recordings used inthis study. This EEG epoch contains the activity of theseizure onset and this epileptic theta activity is mainlypresent on the T2, F8 and T4 electrodes. The seizureEEG is contaminated with EOG and muscle artifacts.Fig. 1 (b) shows the same ictal EEG recording afterEOG artifact correction by the proposed method. Noticethat the EOG artifact was removed, while preserving thesharp quality ictal theta activity on the T2, F8 and T4electrodes. A detailed analysis of the corrected record-ings showed that virtually all blinking artifacts and mostslow eye movements were removed. Fig. 2 shows anothercorrection example from one of the analyzed files. Moresnapshots from every corrected recording as well as theMatlab code used to perform the correction are availablein the internet [13].

0 1 2 3 4 5 6 7 8 9 10

T1−G19

T2−G19

P3−G19

C3−G19

F3−G19

O1−G19

T5−G19

T3−G19

F7−G19

FP1−G19

PZ−G19

CZ−G19

FZ−G19

P4−G19

C4−G19

F4−G19

O2−G19

T6−G19

T4−G19

F8−G19

FP2−G19

Time (seconds)

100 µ V+−

Scale

(a)

0 1 2 3 4 5 6 7 8 9 10

T1−G19

T2−G19

P3−G19

C3−G19

F3−G19

O1−G19

T5−G19

T3−G19

F7−G19

FP1−G19

PZ−G19

CZ−G19

FZ−G19

P4−G19

C4−G19

F4−G19

O2−G19

T6−G19

T4−G19

F8−G19

FP2−G19

Time (seconds)

100 µV+−

Scale

(b)

Fig. 1. (a) Original 10 seconds EEG epoch; (b) EEGrecording after EOG artifact correction. Notice that theEOG artifact was removed completely, while preservingthe sharp quality ictal theta activity in the right tempo-ral lobe (T2 , F8, and lesser degree T4 ).

4. CONCLUSIONS

The proposed automatic EOG correction technique wasapplied on 6 long-term EEG recordings containing ictalactivity. In all cases, most EOG artifacts were removedwithout altering the recorded ictal activity. Therefore,the presented approach is expected to be highly bene-ficial for increasing the accuracy of seizure zone local-ization and for reducing the false positive rates in auto-matic epileptic seizure detection systems. Although themethod does not require any EOG reference channel, wehave observed that the accuracy of the separation pro-duced by SOBI, and therefore of the final correction, isclearly improved when EOG channels are included in theestimation of the hidden source signals. Thus, a topic

132

0 1 2 3 4 5 6 7 8 9 10

T1−G19

T2−G19

P3−G19

C3−G19

F3−G19

O1−G19

T5−G19

T3−G19

F7−G19

FP1−G19

PZ−G19

CZ−G19

FZ−G19

P4−G19

C4−G19

F4−G19

O2−G19

T6−G19

T4−G19

F8−G19

FP2−G19

Time (seconds)

100 µV+−

Scale

(a)

0 1 2 3 4 5 6 7 8 9 10

T1−G19

T2−G19

P3−G19

C3−G19

F3−G19

O1−G19

T5−G19

T3−G19

F7−G19

FP1−G19

PZ−G19

CZ−G19

FZ−G19

P4−G19

C4−G19

F4−G19

O2−G19

T6−G19

T4−G19

F8−G19

FP2−G19

Time (seconds)

100 µV+−

Scale

(b)

Fig. 2. (a) Original 10 s EEG epoch; (b) corrected EEGepoch.

for further research is to compare the performance of theproposed approach with classical automatic correctiontechniques on a dataset with peri-ocular EOG channels.

5. ACKNOWLEDGEMENT

This study was financially supported by the EU projectNo 508803, Biopattern Network of Excellence, by theAcademy of Finland, project No 44876 (Finnish Centreof Excellence Program, 2000-2005), by the FWO throughproject G.0360.05 (EEG signal analysis for epilepsy mon-itoring) and by the Research Council of the KatholiekeUniversiteit Leuven (GOA-AMBioRICS).

6. REFERENCES

[1] R. J. Croft and R. J. Barry, “Removal of ocular artifact fromthe EEG: a review,” Neurophysiol. Clin., vol. 49, pp. 5–19,2000.

[2] Garrick L. Wallstrom, Robert E. Kass, Anita Miller, Jeffrey F.Cohn, and Nathan A. Fox, “Automatic correction of ocularartifacts in the EEG: a comparison of regression-based andcomponent-based methods,” Int. J. of Psychophysiol., vol.53, pp. 105–119, 2004.

[3] S. Puthusserypady and T. Ratnarajah, “H∞ adaptive fil-ters for eye blink artifact minimization from electroencephalo-gram,” IEEE Sig. Proc. Lett., vol. 12, no. 12, pp. 816–819,2005.

[4] T. D. Lagerlund, F. W. Sharbrough, and N. E. Busacker,“Spatial filtering of multichannel electroencephalographicrecordings through principal component analysis by singularvalue decomposition,” J. Clin. Neurophysiol., vol. 14, pp. 73–82, 1997.

[5] P. K. Sadasivan and D. N. Dutt, “SVD based techniquefor noise reduction in electroencephalographi signals,” Sig-nal Processing, vol. 55, pp. 179–189, 1996.

[6] T-P. Jung, S. Makeig, C. Humphries, T-W. Lee, M. J. McK-eown, V. Iragui, and T. J. Sejnowski, “Removing electroen-cephalographic artifacts by blind source separation,” Psy-chophysiology, vol. 37, pp. 163–178, 2000.

[7] A. C. Tang, M. T. Sutherland, and C. J. McKinney, “Valida-tion of SOBI components from high-density EEG,” NeuroIm-age, vol. 25, pp. 539–553, 2004.

[8] C. A. Joyce, I. F. Gorodnitsky, and M. Kutas, “Automaticremoval of eye movement and blink artifacts from eeg datausing blind component separation,” Psychophysiology, vol.41, pp. 313–325, 2004.

[9] S. Faul, L. Marname, G. Lightbody, G. Boylan, and S. Con-nolly, “A method for the blind separation of sources for use asthe first stage of a neonatal seizure detection system,” in Pro-ceedings of the IEEE International Conference on Acoustics,Speech and Signal Processing, 2005, vol. 5, pp. 409–412.

[10] A. Belouchrani, K. Abed-Meraim, J.-F. Cardoso, andE. Moulines, “A blind source separation technique usingsecond-order statistics,” IEEE Transactions on Signal Pro-cessing, vol. 45, pp. 434–444, 1997.

[11] A. Accardo, M. Affinito, M. Carrozzi, and F. Bouquet, “Useof the fractal dimension for the analysis of electroencephalo-graphic time series,” Biol. Cybern., vol. 77, pp. 339–350, 1997.

[12] C. Sevcik, “A procedure to estimate the fractal dimension ofwaveforms,” Complexity International, vol. 5, 1998, Online:http://journal-ci.csse.monash.edu.au/ci/vol05/sevcik/.

[13] “Automatic artifact removal toolbox for Matlab,” 2006,http://www.cs.tut.fi/~gomezher/software/.

133