Clinical Neurophysiology 111 (2000) 1745±1758

www.elsevier.com/locate/clinph

Removal of eye activity artifacts from visual event-related potentials in

normal and clinical subjects
Tzyy-Ping Jung a,b,*, Scott Makeig a,c, Marissa Wester®eld a,d, Jeanne Townsend a,d,
Eric Courchesne a,d, Terrence J. Sejnowski a,b
a
University of California, San Diego, La Jolla, CA 92093, USA
b
Howard Hughes Medical Institute and Computational Neurobiology Lab, The Salk Institute, P.O. Box 85800, San Diego, CA 92186-5800, USA
c
Naval Health Research Center, P.O. Box 85122, San Diego, CA 92186-5122, USA
d
Children's Hospital Research Center, La Jolla, CA 92037, USA
Accepted 23 June 2000

Abstract
Objectives: Electrical potentials produced by blinks and eye movements present serious problems for electroencephalographic (EEG) and
event-related potential (ERP) data interpretation and analysis, particularly for analysis of data from some clinical populations. Often, all
epochs contaminated by large eye artifacts are rejected as unusable, though this may prove unacceptable when blinks and eye movements
occur frequently.
Methods: Frontal channels are often used as reference signals to regress out eye artifacts, but inevitably portions of relevant EEG signals
also appearing in EOG channels are thereby eliminated or mixed into other scalp channels. A generally applicable adaptive method for
removing artifacts from EEG records based on blind source separation by independent component analysis (ICA) (Neural Computation 7
(1995) 1129; Neural Computation 10(8) (1998) 2103; Neural Computation 11(2) (1999) 606) overcomes these limitations.
Results: Results on EEG data collected from 28 normal controls and 22 clinical subjects performing a visual selective attention task show
that ICA can be used to effectively detect, separate and remove ocular artifacts from even strongly contaminated EEG recordings. The results
compare favorably to those obtained using rejection or regression methods.
Conclusions: The ICA method can preserve ERP contributions from all of the recorded trials and all the recorded data channels, even
when none of the single trials are artifact-free. q 2000 Elsevier Science Ireland Ltd. All rights reserved.
Keywords: Artifact removal; Electrooculographic; Independent component analysis; Single-trial event-related potentials; Event-related potential

1. Introduction muscles are often discarded since these are more heavily
contaminated by artifacts than central scalp channels.
Single-trial event-related potentials (ERPs) consist of Another common strategy is to reject all EEG epochs
brief epochs of electroencephalographic (EEG) activity containing artifacts larger than some arbitrarily selected
time-locked to experimental events of interest. These EEG voltage value. However, when limited data are avail-
recordings are usually averaged prior to analysis to increase able, or when blinks and muscle movements occur too
their signal/noise ratio. Here, `noise' includes non-phase- frequently as in children and some patient groups, the
locked EEG signals and non-neural artifacts such as eye amount of data lost to artifact rejection may be unaccepta-
blinks and eye movements. However, ERP averaging may ble. For example, Small (1971) reported a visual ERP
not cancel some artifacts induced by eye movements or experiment conducted on autistic children who produced
blinks if they are time-locked to experimental events. electrooculographic (EOG) artifacts in nearly 100% of the
These artifacts may seriously interfere with correct ERP trials. In this case, the presence of large background EEG
analysis and interpretation. In addition, data from frontal signals not time- and phase-locked to experimental events
and temporal electrodes located near the eyes or scalp may make ERP averages of the few artifact-free trials too
unstable to permit useful analysis.
* Corresponding author. Institute for Neural Computation, University of One approach to reducing contamination from eye move-
California, San Diego, Department 0523, La Jolla, CA 92093-0523, USA. ment artifacts is to regress out reference signals collected
Tel.: 11-619-453-4100, ext. 1455; fax: 11-619-587-0417. near the eyes. Regression methods have been proposed
E-mail address: [email protected] (T.-P. Jung).

1388-2457/00/$ - see front matter q 2000 Elsevier Science Ireland Ltd. All rights reserved. CLINPH 99104
PII: S 1388-245 7(00)00386-2
1746 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

using both time domain (Hillyard and Galambos, 1970; the ICA-based method can also be used to remove stimulus-
Verleger et al., 1982) and frequency domain techniques induced eye artifacts from single-trial ERP records. The
(Whitton et al., 1978; Woestenburg et al., 1983). All regres- method uses spatial ®lters derived by the ICA algorithm,
sion methods, whether in time or frequency domains, avoiding the need for separate reference channels for each
depend on having one or more clean reference channels artifact source, and allowing analysis of neural ERP activity
(e.g. one or more `EOG' channels) which cannot be further projecting to periocular (EOG) channels. Here, we analyze
analyzed after regression. However, these methods share an experimental data collected from 28 normal controls and
inherent weakness, in that both eye movements and EEG from 22 clinical subjects (10 autistic and 12 brain lesion
signals propagate to periocular (`EOG') sites. Therefore, subjects) who had dif®culty in inhibiting unwanted eye
regression-based artifact removal procedures also eliminate movements toward peripheral target stimuli.
neural activity common to the reference electrodes and to
other frontal electrodes. Because the regression coef®cients
are determined largely by the typically large EOG signals, 2. Materials and methods
regression methods may also introduce neural activity
projecting to the reference channel into other sites (Jung 2.1. Subjects
et al., 2000).
Principal component analysis (PCA) has been proposed Data were collected from 28 normal controls, 10 high-
as a method to remove eye artifacts from multichannel EEG functioning autistic and 12 brain lesion subjects. All
(Berg and Scherg, 1991). However, PCA cannot completely subjects had normal or corrected-to-normal vision. The
separate eye artifacts from brain signals, especially when control subjects had no history of substance abuse, special
they have comparable amplitudes (Lagerlund et al., 1997; education, major medical or psychiatric illness, develop-
Jung et al., 1998b, 2000). By combining PCA, multiple mental or neurologic disorder. The autistic subjects met
source models for EOG and EEG, and an artifact-aligned DSM-III-R (American Psychiatric Association, 1987)
averaging method (Lins et al., 1993), Berg and Scherg criteria for autistic disorder, as well as criteria from the
(1994) demonstrated a more effective PCA-based approach Autism Diagnostic Interview, the Autism Diagnostic Obser-
to correct eye artifacts. However, the accuracy of their vation Schedule (Lord et al., 1994) and the Childhood
method depends on the availability of separate and accurate Autism Rating Scale (Schopler et al., 1980). Lesion sites
inverse source solutions for EEG and EOG. Building an in the patients were veri®ed by neuroradiological examina-
accurate EEG source model requires a priori knowledge tion of magnetic resonance images. The stroke and autistic
of event-related brain activity following not only stimuli, subjects had no additional neurologic or psychiatric diag-
but also blinks and saccades. The method also relies on the noses.
amount and quality of separately recorded calibration data
which are needed to provide estimates of the source vectors 2.2. EEG data collection
and transmission coef®cients in the EOG model.
Makeig et al. (1996) proposed an approach to the analysis EEG data were recorded from 31 scalp electrodes, 29
of EEG and ERP data based on an unsupervised neural placed at locations based on a modi®ed International 10±
network learning algorithm that takes a logistic infomax 20 system, one placed below the right eye (VEOG) and one
approach to performing independent component analysis placed at the left outer canthus (HEOG). All 31 channels
(ICA) (Bell and Sejnowski, 1995). They showed that this were referred to the right mastoid and were digitally
ICA algorithm can be used to separate neural activity from sampled for analysis at 256 Hz. Subjects participated in a
muscle and blink artifacts in spontaneous EEG data and 2 h visual spatial selective attention task in which they were
reported its use for ®nding independent components of instructed to attend to ®lled circles ¯ashed in random order
EEG and ERP data and for tracking changes in alertness in 5 boxes laterally arrayed 0.8 cm above a central ®xation
(Makeig et al., 1997; Jung et al., 1998c). Subsequent inde- point. Stimulus locations were outlined by 5 evenly spaced
pendent work (VigaÂrio, 1997) based on a related approach 1.6 cm blue squares displayed on a black background at
veri®ed that different artifacts can also be detected in multi- visual angles of 08, ^2.78 and ^5.58 from ®xation. One
channel magnetoencephalographic (MEG) recordings. (attended) location was marked by a green square through-
However, this study did not attempt to remove the identi®ed out each 72 s experimental block. Subjects were instructed
artifacts. Jung et al. (1998a,b, 2000) introduced an ICA- to maintain ®xation on the central cross and to press a
based method based on an extended infomax ICA algorithm response button as quickly as possible each time they saw
(Girolami, 1998; Lee et al., 1999). This method can be used a ®lled circle appear in the attended location. The location
to detect and remove a wide variety of artifacts (including of the attended square was counterbalanced across experi-
eye blinks, muscle noise, heart signal, and line noise) from mental blocks (for further details see Makeig et al., 1999a;
spontaneous EEG data. Townsend and Courchesne, 1994). Prior to analysis, we
This study demonstrates, through analysis of sample data rejected those trials in which the subject blinked or moved
sets collected in a visual spatial selective attention task, that their eyes at the moment the visual stimulus was presented
 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758 1747

since the eye movement might have impaired the perception details regarding ICA assumptions underlying EEG analy-
of, and response to, the stimulus. sis, see Makeig et al. (1997, 1999a) and Jung et al. (1998a).
Fig. 1 presents a schematic illustration of the procedure.
2.3. Independent component analysis In EEG analysis, the rows of the input matrix, x, are EEG
signals recorded at different electrodes and the columns are
Independent component analysis (ICA) (as de®ned by measurements recorded at different time points (Fig. 1A,
Comon, 1994) is a method for solving the blind source left). ICA ®nds an `unmixing' matrix, W, which decom-
separation problem: to recover N independent source poses or linearly unmixes the multichannel scalp data into
signals, s ˆ {s1 …t†; s2 …t†; ¼; sN …t†} (e.g. different voices, a sum of temporally independent and spatially ®xed compo-
music, or other sound sources) from N linear mixtures, nents, u ˆ Wx. The rows of the output data matrix, u, are
x ˆ {x1 …t†; ¼; xN …t†}, modeled as the result of multiplying time courses of activation of the ICA components. The
the matrix of source activity waveforms, s, by an unknown columns of the inverse matrix, W 21, give the relative
square matrix A (i.e. x ˆ As). Given almost no advance projection strengths of the respective components at each
knowledge of the nature of the sources or of the mixing of the scalp sensors (Fig. 1A, right). These scalp weights
process, the task is to recover a version, u, of the original give the scalp topography of each component, and provide
sources, identical to s, save for scaling and source order. To evidence for the components' physiological origins (e.g. eye
do this, it is necessary to ®nd a square matrix, W, specifying activity should project mainly to far frontal sites). The
®lters that linearly invert the mixing process, i.e. u ˆ Wx. projection of the ith independent component onto the origi-
The key assumption used to distinguish sources from nal data channels is given by the outer product of the ith row
mixtures is that sources, si, are statistically independent, of the component activation with the ith column of the
while their mixtures, xi, are not. In contrast with decorrela- inverse matrix, and is in the original units (e.g. mV). Scaling
tion techniques such as PCA, which ensure only that output information and polarity are distributed between the activa-
pairs are uncorrelated (kui uj l ˆ 0, ; ij), ICA imposes a tion waveforms and the maps, and the true size of a compo-
much stronger criterion, statistical independence, which nent is given only by the size of its projection. In this paper,
occurs when the multivariate probability density function all scalp maps were interpolated from 31 EEG channels and
(p.d.f.) factorizes: e.g. referred to the original right-mastoid reference. For each
Y
N component, the amplitudes of scalp maps (individually
fu …u† ˆ fui …ui † scaled color bars, Fig. 1A, right panels) were the projection
iˆ1 of the component (in mV) at a given time point (vertical blue
Statistical independence requires that all second-order and line).
higher-order correlations of the ui are zero, while decorrela- Artifact-free event-related brain signals were obtained by
tion only seeks to minimize second-order statistics (covar- projecting selected non-artifactual ICA components back
iance or correlation). onto the scalp, x0 ˆ W21 u0 , where u0 is the matrix, u, of
Bell and Sejnowski (1995) proposed a simple neural activation waveforms with rows representing activations of
network `infomax' algorithm that blindly separates artifactual (or irrelevant) sources set to zero (Fig. 1B).
mixtures, x, of independent sources, s, using information
maximization (infomax). They showed that maximizing 2.5. Numerical methods
the joint entropy, H(y), of the output of a neural processor
minimizes the mutual information among the output compo- Extended-ICA decomposition was performed on 31 chan-
nents, yi ˆ g…ui †, where g(ui) is an invertible bounded non- nel, 1 s data epochs from 500 to 700 target stimulus trials for
linearity and u ˆ Wx. Recently, Lee et al. (1999) extended each of 28 normal controls and the 22 clinical subjects using
the ability of the infomax algorithm to perform blind source routines coded in MATLAB 5 and C (available from http://
separation on linear mixtures of sources having either sub- www.cnl.salk.edu/~scott/ica.html) running on a Pentium II
or super-Gaussian distributions based on ideas from Giro- 400 MHz PC with 512 MB RAM. Only target trials in which
lami (1998) and Lee et al. (1999). For further details, see the subject pressed the response button within the allowed
these sources and Jung et al. (2000). (200±1000 ms) time window were analyzed. Previous
results (Makeig et al., 1997, 1999a,b; Jung et al., 2000)
2.4. Applying ICA to single-trial EEG data and those we report here show that ICA decomposition is
relatively insensitive to the exact choice of learning rate or
Use of ICA for blind source separation of EEG data is batch size.
based on two plausible premises: (1) EEG data recorded at Here, we decomposed all the target epochs at once to
multiple scalp sensors are linear sums of temporally inde- study the ICA decomposition of the small evoked responses
pendent components arising from spatially ®xed, distinct or themselves (Makeig et al., 1999a; Jung et al., 1999, 2000).
overlapping brain or extra-brain networks; (2) the spatial However, if artifact removal is the ultimate goal, it is not
spread of electric current from sources by volume conduc- necessary to decompose all 500±700 1 s epochs recorded
tion does not involve signi®cant time delays. For further over 2 h in an experiment at once. Earlier (Jung et al.,
1748 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

Fig. 1. Schematic overview of ICA applied to EEG data. (A) A matrix of single-trial EEG data, x, recorded at multiple scalp sites (only 4 are shown), is used to
train an `unmixing' weight matrix, W, to minimize the statistical dependence of the equal number of outputs, u ˆ Wx (4 shown here). After training, ICA
components consist of time series (the rows of u) giving the time courses of activation of each component, plus ®xed scalp topographies (the columns of W 21)
giving the projections of each component onto the scalp sensors. (B) Some ICA components account exclusively or predominantly for artifactual activity, for
example component IC1, generated by blinks, or IC4, generated by temporal muscle activity. Others account for various evoked and/or spontaneous EEG
activity (e.g. IC2, IC3). Artifact-free EEG signals, x0, can be obtained by mixing and projecting back onto the scalp channels selected non-artifactual ICA
components (IC2 1 IC3) by multiplying the selected activation waveforms, u0, by the inverse mixing matrix, W 21.

1998a,b, 2000), we showed that eye, muscle and line-noise lag regression model of Kenemans et al. (1991). In this
artifacts may be separated from brain activity by training model, the effect of the EOG channel signal on the other
ICA on as little as 10 s of EEG recording, taking less than 1 EEG channels at each sampling time t is given by
min to converge.
X
T
eeg…t† ˆ EEG…t† 2 …bg eog…t 2 g††; where bg ˆ SS21 spg
2.6. Regression analysis gˆ0

To compare the relative effectiveness of ICA for artifact Here EEG denotes the `true' EEG (without eye artifacts),
removal to previous methods, we implemented the multiple- while eeg and eog are the recorded EEG and EOG signals,
 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758 1749

respectively, and T is the maximum time lag. The sequence For each subject, ICA decomposition was performed on
of lagged regression coef®cients, b g, describes the instanta- 500±700 (31 channel, 1 s) data epochs time-locked to target
neous and delayed effects of the EOG on the EEG. The stimulus presentations. We ®rst identi®ed components
vector spg of length (T 1 1) contains the inner products of accounting for blink and eye movement artifacts according
eeg(t) and eog(t 2 g) (g ˆ 0; ¼; T), while SS is the to the procedures described above. The remaining derived
(T 1 1) £ (T 1 1) matrix of inner products of eog(t 2 g). independent components displayed a variety of distinct rela-
This method adjusts for both frequency- and phase-depen- tions to task events. The activations of some components
dent differences in EOG-to-EEG transfer functions (Kene- were clearly time- and phase-locked to stimulus onsets,
mans et al., 1991). while others were time- and phase-locked to button presses
(Jung et al., 1999). Still other components captured various
2.7. Identifying artifactual components types of oscillatory or other background EEG phenomena;
these will be detailed elsewhere.
Most of the 31 independent components derived by ICA
applied to 31 channel single-trial target response data
accounted for distinct activities arising from different 3.1. Removing blink and eye movement artifacts
brain or extra-brain networks. We found blink-related
Example 1. ICA was applied to a target response data set
component(s) among them by the following procedure. (1)
collected from a 32-year-old male autistic subject perform-
We ®rst viewed the 31 channel ICA activations on a scrol-
ing the visual selective attention experiment. Despite our
ling display and searched for the component(s) with time
request that he minimize blinks, about 50% of the trials
courses resembling blink activity (brief, large monopolar
were contaminated by eye blinks as judged by the common
potentials). (2) We veri®ed the nature of the candidate
convention of detecting and rejecting trials of EEG voltages
components by plotting their scalp topographies, which
at periocular sites that exceeded a pre-set threshold.
provided further evidence as to their physiological origin
Extended-ICA successfully isolated blink artifacts to a
(using the heuristic that eye activity projects most strongly
single independent component (Fig. 2A) whose contribu-
to far frontal sites).
tions could be removed from the EEG records by subtract-
Other independent components accounting for eye move-
ing its component projection from the data (Jung et al.,
ments were found using the following procedure. (1) We
1998a,b). Though the subject was instructed to ®xate the
separately averaged the single-trial EEG records time-
central cross during each 72 s block, the technician watch-
locked to target stimuli presented at 5 different locations.
ing the video monitor noticed that the subject's eyes also
(2) We applied the spatial ®lters derived by ICA based on all
tended to move slightly towards target stimuli presented at
the single-trial EEG data to the 5 resulting 31 channel ERPs
peripheral locations. A second ICA component accounted
and searched for components whose time courses varied
for these small horizontal eye movements (Fig. 2B).
systematically with stimulus locations. Our assumption
Fig. 2B (5 traces) shows separate ERP averages (at perio-
here was that involuntary saccadic movements following
cular site HEOG) of responses to targets presented at the 5
stimulus presentations would systematically vary in ampli-
different visual ®eld locations. The size of the prominent
tude and direction as a function of the distance and direction
eye movement-related component is proportional to the
from the ®xation point to the target location. We found such
angle between the stimulus location and the ®xation point.
a systematic relationship for components whose scalp maps
Its scalp pattern is also consistent with the scalp pattern
suggested they accounted for saccadic eye movements, but
expected for lateral eye movements. Note the overlap in
not for other components with different scalp maps account-
scalp topography between the two independent components
ing for other brain (or extra-brain) activities. (3) We veri®ed
accounting for blinks (Fig. 2A) and for lateral eye move-
the nature of the candidate eye movement components by
ments (Fig. 2B). ICA component maps need not be ortho-
examining their scalp topographies. Again, we reasoned that
gonal and may even be nearly spatially coincident.
components accounting for lateral eye movements should
The left panel in Fig. 2C shows all 641 single-trial ERP
project most strongly to far frontal sites, and should show a
epochs recorded at HEOG. Here, we use a recently devel-
polarity difference between the two periocular sites.
oped visualization tool, the `ERP image' (Jung et al., 1999,
2000), to show all 641 single-trial ERP epochs time-locked
3. Results to onsets of target stimuli (left vertical line). Single-trial
event-related responses are plotted as gray-scaled horizontal
We present here the analysis of representative data from traces (see scale bar) sorted by the subject reaction time
two normal (30- and 31-year-old) male control subjects, one (thick black line). The ERP average of these trials is plotted
32-year-old autistic subject and one 55-year-old female below the ERP image. The middle panel in Fig. 2C shows
stroke patient whose lesion involved the right frontal- the summed signals in each trial identi®ed as blink and eye
temporal-parietal cerebral cortex. Results for the remaining movement artifacts. The right panel in Fig. 2C shows the
46 subjects can be seen at http://www.cnl.salk.edu/~jung/ corrected single-trial records obtained by subtracting the
ERPartifact.html. components accounting for artifacts (middle panel) from
1750 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

Fig. 2. Removal of eye blink and eye movement artifacts in normal and autistic subjects. (A) Scalp topography and 5 consecutive 1 s target response epochs
showing the time course of activation of an ICA component accounting for blink artifacts. This component was separated by ICA from 641 target response
trials recorded from an adult autistic subject in the visual selective attention experiment. (B) The scalp topography of a second component (an eye movement
component) and its averaged activation time courses in response to target stimuli presented at the 5 different attended locations. (C) (Left) ERP images of
single-trial ERPs at periocular site HEOG (same autistic subject) time-locked to 641 targets presented at all 5 attended locations, sorted by response time (thick
black line). (Middle) Summed projections at HEOG of the two ICA components accounting for artifacts. (Right) Corrected single-trial ERPs obtained by
subtracting the ICA-extracted artifacts (right) from the original data (left). (D) Averaged ERPs at site HEOG to targets presented at each of 5 attended locations,
before (faint traces) and after (solid traces) artifact removal. Removing blink and eye movement artifacts from the single-trial ERPs revealed the relative
independence of the remaining small visual response from stimulus location. Note that the differences in the target ERPs before removal of the two ICA
components accounting primarily for eye movements are progressively larger in response to stimuli presented away from the central ®xation point, consistent
with a tendency for this subject to move his eyes towards the stimulus locations. (E) The same ICA-based artifact removal procedure applied to single-trial data
collected from a normal control subject. Here, the differences between the uncorrected and corrected ERPs were progressively larger in responses to stimuli
presented away from the central ®xation point.
 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758 1751

the original records (left panel). A large number of blink and ing moderately from heavily contaminated trials ranged
eye movement artifacts (center panel, horizontal line from 70 to 100 mV.
segments) were removed from the records by this procedure. We then applied the spatial ®lters derived by ICA to the
While most ERP reports ignore data collected at perio- data for each subject and removed activities accounting for
cular locations, ICA allows ERP data at periocular electrode artifacts as described in the previous section. Our underly-
sites to be separated into components generated by eye ing assumptions were as follows. (1) If the signals removed
activity and components generated by brain activity. Fig. using ICA did not contain any cerebral activity, the average
2D shows the averaged ERPs from a patient at site HEOG of the least contaminated trials should differ little before and
in response to stimuli presented at the 5 different attended after artifact removal. The average of the heavily contami-
locations before (faint traces) and after (solid traces) artifact nated trials, on the other hand, should show signi®cant
removal. The artifact-corrected averaged periocular differences. (2) If all the eye movement artifacts were effec-
responses to stimuli presented at the 5 different locations tively removed from the EEG records, the corrected ERP
revealed a small periocular positivity near 250 ms that did averages should have very similar waveforms and scalp
not depend on stimulus location. maps, since they were collected from the same scalp loca-
Fig. 2E shows the same artifact removal procedure tions from the same subject performing the same task on the
applied to data collected from a normal control subject. same day.
Conventional threshold-based methods for rejecting Example 2. Fig. 3 shows the ERP averages of the least
contaminated EEG epochs could not wholly avoid summing (351 trials), moderately (132 trials), and heavily contami-
ocular artifacts, since in this example saccade amplitudes nated (72 trials) target trials from a normal control subject
were often smaller than typical (50±75 mV) thresholds for before (Fig. 3A) and after (Fig. 3B) artifact removal. Before
rejecting blinks and eye movements. After artifact correc- artifact removal (Fig. 3A), the averaged ERPs differed
tion by ICA, the responses at site HEOG to targets presented largely because of the contamination from blinks and eye
at the 5 different locations (solid traces) were again quite movements. After artifact removal (Fig. 3B), they differed
similar and showed strong inferior frontal activity recently little, presumably because they now contained only electri-
identi®ed with a fronto-parietal early-P300 subcomponent cal activity of the brain. Note also that the averaged ERPs
(P3f) peaking at the moment of motor command in faster- after artifact removal were very similar to the average of
responding subjects (Makeig et al., 1999a). If, alternatively, least contaminated trials, and the moderately and heavily
the periocular data channels (including HEOG) had been contaminated trial averages showed larger pre-removal
used as reference signals to regress out contributions to and post-removal differences, respectively.
signals at adjacent sites (as in many studies, for example Fig. 3C shows the scalp topographies at the P300 peak
Hillyard and Galambos, 1970; Woestenburg et al., 1983), (505 ms) for the 3 trial averages before (top 3 panels) and
cerebral activity expressed in those channels would have after (lower 3 panels) artifact removal. Note that the scalp
been subtracted from every scalp site, and the reference maps for the 3 `corrected' ERPs (lower 3 panels) are
sites themselves would have become silent. remarkably similar and strongly resemble the artifact-free
scalp map of the least contaminated trials (upper left panel).
This result indicates that ICA may preserve all of the
3.2. Verifying the artifactual nature of the removed signals recorded EEG signals even if the single trials and raw
averages are heavily contaminated by blinks and eye move-
After artifactual signals were removed from the original ments. Note that if a simple rejection method were used on
EEG data, we needed to verify that (1) the signals removed these data (e.g. discarding all trials with maximum EOG
by the ICA-based method did not contain cerebral activity, potentials above 75 mV), the averaged ERPs would be the
and (2) the remaining signals contained no or minimal arti- averages of only the least and moderately contaminated
facts. To this end, we ®rst viewed the histogram of the trials, and that these would still contain some artifacts
maximum potentials (positive or negative) at two periocular even after removal of 13% of the trials.
channels (VEOG and HEOG) for each subject, and used
local minima in the histogram as thresholds for splitting 3.3. ICA-based artifact removal versus artifact rejection
all the single target response trials (from 100 ms before to
900 ms after stimulus onsets) into 3 subsets: least contami- Example 3. We used 545 single trials collected from a
nated, moderately contaminated, and most heavily contami- patient with frontal-temporal-parietal lesion to compare the
nated, according to their absolute maximum potential values effectiveness of artifact removal using the ICA-based
at the two periocular channels. Using this method, the method with results of the standard artifact rejection
threshold voltage separating least from moderate and method. The artifact rejection procedure employed in the
moderate from heavily contaminated trials varied between data collection laboratory was to discard blink contaminated
individual subjects. Across all 50 subjects, the threshold trials containing potentials exceeding 75 mV at either sites
separating least contaminated from moderately contami- VEOG or HEOG. This procedure rejected all but 49 of 545
nated trails ranged from 25 to 40 mV; the threshold separat- trials of the patient's data. Since blinks and muscle move-
1752 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

Fig. 3. Removal of eye blink and eye movement artifacts from single trials. (A) Averages of least (red traces, 351 trials), moderately (blue traces, 132 trials) and
heavily (green traces, 72 trials) contaminated single-trial target response epochs from a 31 year-old normal control subject. (B) Averages of artifact-corrected
ERPs for the 3 trial groups. (C) Scalp topographies at the P300 peak (505 ms) for the least (left), moderately (center) and heavily (right) contaminated trial
groups before (top) and after (bottom) artifact removal.

ments occurred so frequently, the amount of data lost to artifact rejection appears highly contaminated by large
standard artifact rejection would be unacceptable, and so alpha-frequency EEG activity. However, at frontal sites its
use of an effective artifact removal procedure was crucial (low-frequency) contour resembles the average of all 545
to the use of this patient's data for further analysis. `ICA-corrected' trials more than that of the uncorrected
Fig. 4A shows the averaged target responses of the lesion trials, supporting the inference that the signals removed by
subject before (blue traces) and after (red traces) ICA-based the artifact removal procedure were indeed artifacts. At
artifact removal and standard artifact rejection (green posterior sites which were relatively uncontaminated by
traces). Our assumption underlying this comparison was eye movements, problems resulting from averaging an
that after an effective artifact removal procedure, the insuf®cient number of `artifact-free' (e.g. 49) trials are
corrected averages would resemble the average of `arti- apparent. The average is overlaid with strong remnants of
fact-free' trials (if enough of these were recorded). alpha EEG activity that clearly compromise accurate
In the ®gure, the average of the 49 remaining trials after measures of ERP peak amplitude and latency. The ICA-
 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758 1753

corrected average is free both of the large eye movement pattern because of alpha contamination, while the corrected
artifacts present in the uncorrected average of all trials, and trial average generally resembled the `P300' scalp topogra-
of the non-phase-locked background EEG present in the phy found for normal control subjects (Makeig et al.,
average of the few `artifact-free' trials. 1999a), albeit with smaller than normal amplitude.
Fig. 4B shows the scalp maps at the `P300' peak (near Fig. 4C shows the averaged ERPs for this patient at site
457 ms) in averages of all 545 uncorrected trials (left), the HEOG in response to target stimuli presented at the 5 differ-
49 `artifact-free' trials (center) and of all 545 trials after ent attended locations, before (blue traces) and after (red
artifact removal (right). As expected, the P300 scalp map traces) ICA-based artifact removal. As has been observed
for the uncorrected trial average included a strong negativity in patients with similar lesions (HeÂcaen, 1962; LaÂdavas et
at far frontal sites produced by blinks and eye movements. al., 1997), the patient clearly found it dif®cult to inhibit
The artifact-free average showed no recognizable scalp ipsilesional eye movements, as indicated by the progres-

Fig. 4. ICA correction versus artifact rejection in a patient with right-frontal-temporal-parietal lesion. (A) Averages of responses to target stimuli before (blue
traces) and after (red traces) ICA-based artifact correction. The (red traces) average of all 545 artifact-corrected trials strongly resembles the (green traces)
average of 49 trials resulting from the conventional artifact rejection method, while avoiding the residual EEG `contamination' remaining in the 49 trial
averages because of insuf®cient cancellation of strong non-phase-locked alpha and other EEG processes. (B) Scalp topographies at the P300 peak (457 ms) of
the (left) uncorrected average, (center) artifact-rejected average, and (right) ICA-corrected response average. (C) Averaged ERPs at site HEOG to targets
presented at each of 5 attended locations, before (blue traces) and after (red traces) ICA-based artifact removal. Note that the subject was unable to inhibit her
eye movements towards targets ipsilateral to the lesion. ICA-corrected averages (red traces) show no such lateralized differences.
1754 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

Fig. 5. ICA correction versus regression correction in removing artifacts from ERPs. (A) The raw averages of least (red traces, 397 trials), moderately (blue
traces, 105 trials) and heavily (green traces, 64 trials) contaminated trials from a normal control subject before any artifact correction, (B) after ICA correction,
and (C) after correction by multiple-lag regression. Note that the least contaminated trails were not subjected to regression here. (D) The scalp topographies at
the P300 peak (423 ms) of least (left), moderately (middle) and heavily (right) contaminated trials before (top) and after (middle) artifact correction by ICA and
regression (bottom).

sively larger eye movement artifacts in the uncorrected that the ICA-based artifact removal identi®ed and removed
averaged responses to target stimuli presented to the right these mostly unilateral eye movement artifacts, revealing
of ®xation. The corrected ERP averages (red traces) show underlying neuronal phenomena consistent across stimulus
 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758 1755

locations that would be dif®cult or impossible to measure into separate components with physiologically plausible
accurately using other analysis methods. scalp maps.

3.4. ICA-based artifact removal versus regression 4.1. ICA correction versus artifact rejection

Example 4. This example uses a data set from a normal Rejecting contaminated EEG segments using threshold
control to compare the use of ICA with multiple-lag regres- voltage criteria, the most commonly used method for deal-
sion to correct for eye movement artifacts. Again, we ®rst ing with artifacts in research settings, can be unacceptable
viewed a histogram of the maximal potentials in each epoch when blinks and muscle movements occur too frequently. In
at the two periocular channels and then split all single target contrast, ICA-based artifact removal can effectively detect
response trials into 3 groups: those least (maxumVu # 25), and separate contaminations arising from a wide variety of
moderately (25 , maxumVu # 35), and most heavily artifactual sources in EEG records without losing neural
(maxumVu . 35) contaminated by eye movements. signals recorded at frontal and even periocular sites. This
The regression method employed the recorded signals at method can preserve most or all of the recorded trials for
sites VEOG and HEOG as reference channels. As recom- analysis (Fig. 4, red traces), even when few or none of the
mended, regression was performed only for single-trial single trials are artifact-free. In our example, ICA also
epochs containing maximal absolute potentials greater recovered from the artifact-laden data of clinical subjects
than 25 mV. Fig. 5A shows the averages of the 3 groups the same `P300' scalp topography (Fig. 4B, right panel)
of trials before artifact removal. The averages of trials generally obtained from normal control subjects. Before
`corrected' by ICA and by regression for the least (red, artifact removal, this pattern was overwhelmed by a strong
397 trials), moderately (blue, 105 trials) and heavily artifactual negativity at frontal sites and was thus very dif®-
(green, 64 trials) contaminated trial groups are shown in cult to identify either in the scalp map of the raw data
Fig. 5B,C, respectively. Fig. 5D shows the scalp topogra- averages or the regression-corrected averages (Fig. 4B,
phies at the P300 peak (423 ms) for the 3 ERP averages left and middle panels). Additionally, the result of ICA-
before (top 3 panels) and after (middle 3 panels) artifact based artifact removal illustrates the importance of having
removal by ICA and by regression (bottom 3 panels). The an effective artifact removal method when ERPs from clin-
3 ICA-corrected averages had very similar waveforms and ical populations, whose data may be heavily contaminated
scalp topographies, while the regression-corrected averages by involuntary eye movements or muscle tension, are to be
had smaller potentials for the moderately (blue) and heavily measured and interpreted.
(green) contaminated trials compared to the averages of the
least contaminated (`artifact-free') trials (red traces). After 4.2. ICA correction versus regression correction
removal of eye movement artifacts by regression (middle
trace), the neural signals contributing to the reference The current alternative to artifact rejection is regression
signals (VEOG and HEOG) were also eliminated from the in the time or frequency domain. This is performed using
other frontal scalp sites. Also note that the averages of conceptually distinct sets of EEG and EOG channel data to
moderately and heavily contaminated trials contain no derive parameters characterizing the appearance and spread
signals at sites VEOG and HEOG, whereas the ICA- of EOG artifacts in the EEG channels. However, because
corrected averages reveal the neural activity contributing EEG and ocular activity mix bidirectionally (Peters, 1967;
to the recorded signals at these sites. Oster and Stern, 1980), regressing out eye artifacts inevita-
bly involves subtracting relevant neural signals from each
record as well as ocular activity. Regression methods
4. Discussion become even more problematic when appropriate regressing
channels are not available. For example, regressing out
The goal of the present study was to determine whether muscle artifacts using this method would require a reference
ICA could be used to remove artifacts of non-neural origin channel selective for each contributing muscle. ICA-based
from single ERP data trials particularly in clinical subjects methods, on the other hand, can be used to separate and
that are heavily contaminated with eye movement artifacts, remove multiple muscle artifacts as well as ocular artifacts,
thereby preserving the recorded event-related brain activity. as we have shown elsewhere (Jung et al., 2000).
Here, ICA was applied to single-trial target response records The present results (Fig. 5) showed that regression-based
from a total of 50 (28 normal, 10 autistic and 12 brain artifact removal of moderately and heavily contaminated
lesion) subjects in a visual selective attention experiment. trials consistently overcorrected and removed neural activ-
For each subject, ICA derived spatial ®lters or maps that ity from electrodes located over frontal and periocular sites.
decomposed EEG data recorded at multiple scalp sensors ICA, on the other hand, recovered the EEG activity without
into a sum of components with ®xed scalp distributions and relying on the availability of one or more `uncontaminated'
maximally independent time courses. The spatial ®lters reference channels, even when applied to heavily contami-
clearly separated eye blink and eye movement artifacts nated trials. The resulting corrected response averages were
1756 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

in each case remarkably similar to the averages of the least reason to believe that cerebral and artifactual sources in
contaminated trials (Fig. 5). the spontaneous EEG necessarily remain ®xed over time
or occurrences. Examples of non-®xed sources may include
4.3. ICA correction versus PCA correction spreading sleep spindles (McKeown et al., 1998). However,
in our studies of averaged and unaveraged data from normal
Compared to PCA that minimizes second-order correla-
control subjects in these experiments (Jung et al., 1999), the
tions among the output channels, ICA imposes a much
relatively small numbers of obtained components showing
stronger criterion, statistical independence. Statistical inde-
stimulus-locked, response-locked, and non-phase-locked
pendence requires that all second-order and higher-order
categories, each accounting for activity occurring across
correlations reduce to zero. PCA ®nds orthogonal directions
sets of 500 or more 1 s trials, suggests that the brain areas
of greatest variance in the data, whereas ICA component
generating our data were primarily ®xed. This supposition is
maps could be non-orthogonal. In general, there is no reason
concordant with repeated observations in functional brain
why neurobiologically distinct artifact and EEG sources
imaging experiments that discrete, spatially restricted areas
should be spatially orthogonal to one another. Since ICA
of cortex are activated during task performance (Friston et
does not impose any condition on the spatial ®lters, it can
al., 1998), and has been further veri®ed for these data in a
collect concurrent activity arising from spatially overlap-
computationally intense moving-average ICA study
ping artifact and EEG source distributions. We have found
(Makeig et al., 2000b).
that ICA produces components whose scalp maps have only
a few spatial maxima, consistent with relatively compact or
simply connected source generators. In comparison, PCA 4.5. Advantages of ICA-based artifact removal
produces components that mostly have more complex
Our results show that ICA has at least 3 advantages
spatial patterns (Silberstein and Cadusch, 1992; Jung et
compared with other artifact removal methods. (1) ICA
al., 1998b, 2000).
simultaneously separates EEG signals including artifacts
Elsewhere, we have shown that ICA can effectively
into independent components based on the characteristics
remove contamination from a wide variety of artifactual
of the data, without relying on the availability of one or
sources in spontaneous EEG data with results comparing
more `clean' reference channels for each type of artifact.
favorably to those obtained using regression and PCA meth-
This avoids the problem of mutual contamination between
ods (Jung et al., 1999, 2000). We have also presented
regressing and regressed channels. (2) ICA-based artifact
evidence for the behavioral consistency of ICA decomposi-
removal can preserve all of the recorded trials, a crucial
tion applied to ERP averages (Makeig et al., 1999a,b).
advantage over rejection-based methods when limited data
4.4. ICA limitations are available, or when blinks and muscle movements occur
too frequently, as in some patient groups. (3) Unlike regres-
Although ICA appears to be quite useful for EEG analy- sion methods, ICA-based artifact removal can preserve data
sis, it is important to keep in mind that it also has some at all scalp channels, including frontal and periocular sites.
inherent limitations. ICA uses spatial ®ltering to achieve artifact removal; the
First, infomax ICA can decompose at most N sources ®lters derived from a brief EEG recording portion (e.g. 5±10
from N scalp electrodes. Usually, the effective number of s) might even be applied to multichannel EEG by a simple
temporally-independent signals contributing to the scalp matrix multiplication to successive EEG epochs to create an
EEG is unknown, and it is likely that observed brain activity artifact-reduced EEG in real time. More exhaustive ICA
arises from more physically separable effective sources than decomposition may be needed only when the spatial ®lters
the available number of EEG electrodes. We have explored fail to effectively remove artifacts with very different spatial
and discussed this issue elsewhere (Makeig et al., 2000a). distributions. We thus believe it may be possible to perform
Second, the assumption of temporal independence used artifact removal in routine clinical EEG in near real time
by ICA cannot be satis®ed when the training data set is too using an appropriate hardware and software implementa-
small, or when separate topographically distinguishable tion.
phenomena always occur concurrently in the data. In the In addition to artifact removal, ICA decomposition can be
latter case, simulations show that ICA may derive a compo- highly useful for enhancing the amount and quality of infor-
nent accounting for their joint occurrence, plus separate mation in event- or response-related brain signals that can
components accounting for their periods of solo activation. be extracted from ERP data (Makeig et al., 1997, 1999a),
Such confounds imply that converging behavioral or other and examining systematic variations from trial to trial
evidence must be obtained before concluding that spatio- within subjects (see Jung et al., 1999). The analysis of
temporally overlapping ICA components measure neuro- single-trial data is particularly important in studies of orient-
physiologically or functionally distinct activities. ing, habituation, or associative learning. However, most or
Third, ICA assumes that the physical sources of artifac- all of the information available in single trials is usually
tual and neural activity contributing to EEG signals are sacri®ced to increase signal-to-noise ratio through within-
spatially stationary through time. In general, there is no or between-subject averaging (Kenemans et al., 1989).
 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758 1757

Removing EEG artifacts from brain activity by ICA can Jung T-P, Makeig S, Wester®eld M, Townsend J, Courchesne E, Sejnowski
increase the amount of behaviorally and neurophysiologi- TJ. Analyzing and visualizing single-trial event-related potentials. Adv
Neural Info Processing Systems 1999;11:118±124.
cally relevant information available in ERP data, and may Jung T-P, Makeig S, Humphries C, Lee T-W, McKeown MJ, Iragui V,
also give researchers the ability to examine trial to trial Sejnowski TJ. Removing electroencephalographic artifacts by blind
response variations. Similar ICA approaches should be source separation. Psychophysiology 2000;37:163±178.
equally applicable to other types of multichannel biomedi- Kenemans JL, Verbaten MN, Roelofs JW, Slangen JL. ªInitial-º and
cal data for which linear superposition of the contributing ªchange-orienting reactionsº: an analysis based on visual single-trial
event-related potentials. Biol Psychol 1989;28(3):199±226.
signals can be assumed (e.g. electromyography, magnetoen- Kenemans JL, Molenaar P, Verbaten MN, Slangen JL. Removal of the
cephalography, electrocardiography, etc.). ocular artifact from the EEG: a comparison of time and frequency
domain methods with simulated and real data. Psychophysiology
1991;28(1):114±121.
Acknowledgements LaÂdavas E, Zeloni G, Zaccara G, Gangemi P. Eye movements and orienting
of attention in patients with visual neglect. J Cog Neurosci
This report was supported in part by a grant from the 1997;9(1):67±74.
Swartz Foundation, and grants from the Of®ce of Naval Lagerlund TD, Sharbrough FW, Busacker NE. Spatial ®ltering of multi-
channel electroencephalographic recordings through principal compo-
Research ONR.Reimb.30020.6429 (S.M.), from the
nent analysis by singular value decomposition. J Clin Neurophysiol
Howard Hughes Medical Institute (T.J.S.), and from the 1997;14(1):73±82.
National Institute of Health NIMH 1-R01-NH-36840 and Lee T-W, Girolami M, Sejnowski TJ. Independent component analysis
NINDS 1RO1-NS34155-01 (E.C. and J.T.). The views using an extended infomax algorithm for mixed sub-Gaussian and
expressed in this article are those of the authors and do super-Gaussian sources. Neural Computation 1999;11(2):606±633.
Lins OG, Picton TW, Berg P, Scherg M. Ocular artifacts in recording EEG
not re¯ect the of®cial policy or position of the Department
and event-related potentials II: source dipoles and source components.
of the Navy, Department of Defense, or the US Govern- Brain Topogr 1993;6(1):65±78.
ment. Lord C, Rutter M, Le Couteur A. Autism Diagnostic Interview-Revised: a
revised version of a diagnostic interview for caregivers of individuals
with possible pervasive developmental disorders. J Aut Dev Disord
References 1994;24(5):659±685.
Makeig S, Bell AJ, Jung T-P, Sejnowski TJ. Independent component analy-
American Psychiatric Association. Diagnostic and statistical manual of sis of electroencephalographic data. Adv Neural Info Processing
mental disorders, 3rd ed. revised. Washington, DC: American Psychia- Systems 1996;8:145±151.
tric Association, 1987. Makeig S, Jung T-P, Bell AJ, Ghahremani D, Sejnowski TJ. Blind separa-
Bell AJ, Sejnowski TJ. An information-maximization approach to blind tion of event-related brain responses into independent components.
separation and blind deconvolution. Neural Computation Proc Natl Acad Sci USA 1997;94:10979±10984.
1995;7:1129±1159. Makeig S, Wester®eld M, Jung T-P, Covington J, Townsend J, Sejnowski
Berg P, Scherg M. Dipole models of eye activity and its application to the TJ, Courchesne E. Independent components of the late positive
removal of eye artifacts from the EEG ad MEG. Clin Physiol Meas response complex in a visual spatial attention task. J Neurosci
1991;12(Suppl A):49±54. 1999a;19(7):2665±2680.
Berg P, Scherg M. A multiple source approach to the correction of eye Makeig S, Wester®eld M, Townsend J, Jung T-P, Courchesne E, Sejnowski
artifacts. Electroenceph clin Neurophysiol 1994;90(3):229±241. TJ. Independent components of early event-related potentials in a visual
Comon P. Independent component analysis ± a new concept? Signal spatial attention task. Phil Trans Biol Sci 1999b;354:1135±1144.
Processing 1994;36(3):287±314. Makeig S, Jung T-P, Ghahremani D, Sejnowski TJ. Independent component
Friston KJ, Fletcher P, Josephs O, Holmes A, Rugg MD, Turner R. Event- analysis of simulated ERP data. In: Nakada T, editor. Human higher
related fMRI: characterizing differential responses. Neuroimage function I: advanced methodologies, 2000a, in press.
1998;7(1):30±40. Makeig S, Enghoff S, Jung T-P, Sejnowski TJ. Moving-window ICA
Girolami M. An alternative perspective on adaptive independent compo- decomposition of EEG data reveals event-related changes in oscillatory
nent analysis algorithms. Neural Computation 1998;10(8):2103± brain activity. The 2nd International Workshop on Independent Compo-
2114. nent Analysis and Signal Separation, 2000b, in press.
HeÂcaen H. Clinical symptomatology in right and left hemisphere lesions. McKeown MJ, Humphries C, Achermann P, Borbely AA, Sejnowski TJ. A
In: Mountcastle VB, editor. Interhemispheric relation and cerebral new method for detecting state changes in the EEG: exploratory appli-
dominance, Baltimore, MD: John Hopkins University Press, 1962. cation to sleep data. J Sleep Res 1998;7(Suppl 1):48±56.
Hillyard SA, Galambos R. Eye-movement artifact in the CNV. Electroen- Oster PJ, Stern JA. Measurement of eye movement electrooculography. In:
ceph clin Neurophysiol 1970;28(2):173±182. Matin I, Venables PH, editors. Techniques in psychophysiology,
Jung T-P, Humphries C, Lee T-W, Makeig S, McKeown MJ, Iragui V, Chichester: Wiley, 1980. pp. 275±309.
Sejnowski TJ. Extended ICA removes artifacts from electroencephalo- Peters JF. Surface electrical ®elds generated by eye movement and eye
graphic data. Adv Neural Info Processing Systems 1998a;10:894± blink potentials over the scalp. J EEG Technol 1967;7:27±40.
900. Schopler E, Reichler RJ, De Velis RF, Daly K. Toward objective classi®ca-
Jung T-P, Humphries C, Lee T-W, Makeig S, McKeown MJ, Iragui V, tion of childhood autism: childhood autism rating scale (CARS). J Aut
Sejnowski TJ. Removing electroencephalographic artifacts: compari- Dev Disord 1980;10:91±103.
son between ICA and PCA. Neural Networks Signal Processing Silberstein RB, Cadusch PJ. Measurement processes and spatial principal
1998b;VIII:63±72. components analysis. Brain Topogr 1992;4(4):267±276.
Jung T-P, Makeig S, Bell AJ, Sejnowski TJ. Independent component analy- Small JG. Sensory evoked responses of autistic children. Infantile Autism
sis of electroencephalographic and event-related potential data. In: 1971:224±239.
Poon P, Brugge J, editors. Auditory processing and neural modeling, Townsend J, Courchesne E. Parietal damage and narrow ªspotlightº spatial
New York: Plenum Press, 1998c. pp. 189±197. attention. J Cog Neurosci 1994;6:220±232.
1758 T.-P. Jung et al. / Clinical Neurophysiology 111 (2000) 1745±1758

Verleger R, Gasser T, MoÈcks J. Correction of EOG artifacts in event-related Whitton JL, Lue F, Moldofsky H. A spectral method for removing eye-
potentials of EEG: aspects of reliability and validity. Psychophysiology movement artifacts from the EEG. Electroenceph clin Neurophysiol
1982;19(4):472±480. 1978;44:735±741.
VigaÂrio RN. Extraction of ocular artifacts from EEG using independent Woestenburg JC, Verbaten MN, Slangen JL. The removal of the eye-move-
component analysis. Electroenceph clin Neurophysiol 1997;103:395± ment artifact from the EEG by regression analysis in the frequency
404. domain. Biol Psychol 1983;16:127±147.