A Robust Method to Filter Various Types of Artifacts on Long Duration EEG Recordings
ABSTRACT EEG is a system used to measure electrical brain activity using multiple electrodes placed on the scalp. Unfortunately, the signals can be easily contaminated by noises called artifacts. These can be generated by various actions such as eye blinks, eye movements, muscle activities or small electrode movements. This paper presents a global artifact removal method corresponding to an evolution of the AFOP method (Adaptive Filtering by Optimal Projection) in order to improve its stability. This evolution automatically filters ocular, muscular and heart beat artifacts. The results are validated on long duration EEG recordings containing pathological activities. An expert analysis shows that the cerebral signal is well conserved while a lot of artifacts are removed.
[show abstract] [hide abstract]
ABSTRACT: A new approach to filter multi-channel signals is presented, called filtering by optimal projection (FOP) in this paper. This approach is based on common spatial subspace decomposition (CSSD) theory. Moreover, an evolution of this method for non-stationary signals is also introduced which is called adaptative FOP (AFOP). As ICA, a filtering matrix is set up in the best way to remove artifacts with linear combination of channels. This filtering matrix is characterized by two subspaces. The first one is determined during a learning phase, by finding components maximizing the ratio signal over noise. The second one will be determined during a filtering phase, by reconstructing signals of a sliding window, by a least square method. These methods are completely automated and enable to filter independently numerous artifact types. Moreover, this filtering can be improved by applying this process on frequency band decomposed signals.Various tests have been made on electroencephalogram (EEG) signals in order to remove ocular and muscular activity while conserving pathological activity (slow waves, paroxysms). The results are compared with ICA filtering and medical inspection has been carried out to prove that this approach yields very good performance.Signal Processing.
Article: A system for automatic artifact removal in ictal scalp EEG based on independent component analysis and Bayesian classification.[show abstract] [hide abstract]
ABSTRACT: To devise an automated system to remove artifacts from ictal scalp EEG, using independent component analysis (ICA). A Bayesian classifier was used to determine the probability that 2s epochs of seizure segments decomposed by ICA represented EEG activity, as opposed to artifact. The classifier was trained using numerous statistical, spectral, and spatial features. The system's performance was then assessed using separate validation data. The classifier identified epochs representing EEG activity in the validation dataset with a sensitivity of 82.4% and a specificity of 83.3%. An ICA component was considered to represent EEG activity if the sum of the probabilities that its epochs represented EEG exceeded a threshold predetermined using the training data. Otherwise, the component represented artifact. Using this threshold on the validation set, the identification of EEG components was performed with a sensitivity of 87.6% and a specificity of 70.2%. Most misclassified components were a mixture of EEG and artifactual activity. The automated system successfully rejected a good proportion of artifactual components extracted by ICA, while preserving almost all EEG components. The misclassification rate was comparable to the variability observed in human classification. Current ICA methods of artifact removal require a tedious visual classification of the components. The proposed system automates this process and removes simultaneously multiple types of artifacts.Clinical Neurophysiology 05/2006; 117(4):912-27. · 3.41 Impact Factor
[show abstract] [hide abstract]
ABSTRACT: Eye movements, eye blinks, cardiac signals, muscle noise, and line noise present serious problems for electroencephalographic (EEG) interpretation and analysis when rejecting contaminated EEG segments results in an unacceptable data loss. Many methods have been proposed to remove artifacts from EEG recordings, especially those arising from eye movements and blinks. Often regression in the time or frequency domain is performed on parallel EEG and electrooculographic (EOG) recordings to derive parameters characterizing the appearance and spread of EOG artifacts in the EEG channels. Because EEG and ocular activity mix bidirectionally, regressing out eye artifacts inevitably involves subtracting relevant EEG signals from each record as well. Regression methods become even more problematic when a good regressing channel is not available for each artifact source, as in the case of muscle artifacts. Use of principal component analysis (PCA) has been proposed to remove eye artifacts from multichannel EEG. However, PCA cannot completely separate eye artifacts from brain signals, especially when they have comparable amplitudes. Here, we propose a new and generally applicable method for removing a wide variety of artifacts from EEG records based on blind source separation by independent component analysis (ICA). Our results on EEG data collected from normal and autistic subjects show that ICA can effectively detect, separate, and remove contamination from a wide variety of artifactual sources in EEG records with results comparing favorably with those obtained using regression and PCA methods. ICA can also be used to analyze blink-related brain activity.Psychophysiology 02/2000; 37(2):163 - 178. · 3.29 Impact Factor
A robust method to filter various types of artifacts
on long duration EEG recordings
Samuel Boudet∗, Laurent Peyrodie∗, Philippe Gallois‡and Christian Vasseur§
∗HEI-ERASM (Hautes Etudes d’Ing´ enieur) and LAGIS, Lille, 59000, France
Email: email@example.com and firstname.lastname@example.org
‡St Vincent Hospital, Lille, France
§LAGIS UMR-CNRS 8146, University of Lille1, Villeneuve d’Ascq, 59655, France
Abstract—EEG is a system used to measure electrical brain
activity using multiple electrodes placed on the scalp. Unfortu-
nately, the signals can be easily contaminated by noises called
artifacts. These can be generated by various actions such as
eye blinks, eye movements, muscle activities or small electrode
movements. This paper presents a global artifact removal method
corresponding to an evolution of the AFOP method (Adaptive
Filtering by Optimal Projection) in order to improve its stability.
This evolution automatically filters ocular, muscular and heart
beat artifacts. The results are validated on long duration EEG
recordings containing pathological activities. An expert analysis
shows that the cerebral signal is well conserved while a lot of
artifacts are removed.
Artifacts represent a major problem for Electroencephalog-
raphy analysis. Artifacts are defined by signal alterations
caused by non-cerebral activities and can have several origins.
Ocular artifacts appear when the patient blinks or moves his
eyes, muscular artifacts appear when the patient contracts his
jaw or his forehead, and electrode artifacts (including heart
beat artifact) are generated by electrode mechanical movement.
Artifact amplitude is often much higher than EEG ampli-
tude, thus making the analysis difficult and making automatic
processing impossible. Our team has been working on this
subject for years  in order to improve automatic treatment
of EEG for anticipated detection of epilepsy seizures.
This paper deals with the filtering of long duration EEG
recordings (3 to 6 hours) for clinical applications in order
to ease the neurophysiologic interpretation. The long duration
examinations are often more artifacted than standard examina-
tions (20 min) because the patients are not asked to relax. The
signal over noise ratio can then be very low and all studies
show that it is impossible to completely filter high amplitude
artifacts without removing the EEG signal .
This paper presents a global method enabling to filter a large
amount of artifacts while preserving the major part of EEG
activity and particularly pathological activities. This method
is an improvement of the AFOP method (Adaptive Filtering
by Optimal Projection) published by the authors , which
improves its stability. The AFOP method filters ocular and
muscular artifacts but is not well designed to filter electrode
artifacts. This article also proposes a way to filter heart beat
artifacts described in section II, using the idea of Y. Wang .
The medical context is described in the first part in order
to understand what has to be filtered or not. Then, a brief
overview of spatial filtering is introduced by presenting back-
ground on ICA. The principle of AFOP and its improvement
is presented in the third part. Finally, a comparison between
manual ICA and the proposed method is made to prove that
this method can improve filtering in several cases.
II. MEDICAL CONTEXT
The authors have been working on a long duration exam-
inations (3 to 6 hours) recording with a 19 electrodes 10/20
system. Two electrodes are added on each wrist to measure
Electrocardiographic activity. During an examaniation, the
patient is exposed to flashing lights at various frequencies and
also carries out hyperventilation, which consists in breathing
quickly and deeply. Sometimes the patient is also asked to
sleep and may enter in the firsts sleeping phases. These tests
can reveal pathological troubles or cause epilepsy seizures.
During the standard examination, some elements must be
kept due to the fact that they correspond to cerebral activity.
It can be noticed especially:
• Paroxysms (epilepsy indicators). They are graphical el-
ements that can be seen under various forms. The most
common one is called spike wave. The paroxysms be-
longs to a very large frequency band (1 − 30 Hz).
• Reactivity to eye closure. This is characterized by an
alpha rhythm apparition (8 − 13 Hz) mainly located in
the occipital region.
• Slow waves (< 4 Hz). These waves can be seen as
pathological for the adult. However, this activity is normal
for people under 25 years old during hyperpnoea.
The main purpose of the filtering process is to keep these
signals as much as possible and to erase the ones having
artifactual origins. The artifact types can be broken down as
• Ocular artifacts. Two kinds of ocular artifacts can be
observed: eye blinks and eye movements. Eye blinks are
represented by a low frequency signal (< 4 Hz) with
high amplitude. It is a symmetrical activity mainly located
on the front electrodes (FP1, FP2) with low propagation.
Eye movements are also represented by a low frequency
signal (< 4 Hz) but with higher propagation, especially
on the temporal electrodes. It is caused by the fact that
eyes represent dipoles and their movements lead to an
alteration of the electrical field. It is characterized by a
dissymmetry between the two hemispheres.
• Muscle artifacts. Muscular activity creates high fre-
quency signals (> 13 Hz). There are a lot of muscles all
over the head, but the muscles in the forehead and jaw
are strongly visible. Jaw muscles are powerful and can
produce an important signal on temporal area. Forehead
muscles are less powerful but are closer to electrode and
produce a signal on frontal area.
• Mechanical and electrode movements. It can be caused
by movomvement of the wire or by bad connection
between an electrode and the skin. In this case the signal
is only located on one electrode. Mechanical artifacts
may also occur when the patient contracts his muscle.
A movement of the facial skin or of the head part can
then be generated which result in an artifact signal on
several electrodes. Finally, mechanical artifacts can be
caused by heart beats if an electrode is located on a
vein. This will create an activity synchronized with the
heart but having a more sinusoidal shape. Mechanical and
electrode artifacts are mainly low frequency signals (< 4
Hz) due to the fact that they directly correspond to the
III. STATE OF THE ART
For an awake adult, the main part of measured signal is in
the frequency band α (8 - 13 Hz) and the artifacts are either at
higher frequencies (> 13 Hz) for muscle artifacts, or at lower
frequencies (< 8 Hz) for ocular and mechanical artifacts.
However, some EEG signals (particularly pathological element
like slow waves and spikes) can occur on the artifactual
frequency bands and make a frequency filtering impossible.
This is why a lot of work is realized on spatial filters which
use information of signal repartition on various channels. Most
of the methods use Independent Component Analysis (ICA)
to realize this kind of filter .
A. Spatial filter using ICA
ICA supposes that the various signal channels are a linear
mixing of sources :
V = MS
V is the signal matrix where lines represent channels and
columns time samples and S is the signal matrix of sources.
The matrix M(m,n) is called mixing matrix. If the number
of sources (n) is inferior or equal to the number of channels
(m), M will be invertible. The pseudo-inverse matrix W =
(MTM)−1MTis called separating matrix.
ICA aims to estimate this separating matrix W so that
the sources (S = WV) would be independent. There exists
many methods of ICA and each of them uses a different
measure of independence called contrast function . For
example, T. Jung et al. uses Infomax method which maximizes
neguentropy of the sources . Once they are defined, they
are identified as artifactual or not and the artifactual ones are
canceled: S?= DS, where D is a diagonal matrix with 0 on
artifacted components and 1 on the others. The signals are
then reconstructed by inverse transformation.
V?represents filtered signals. A filtering matrix is built using:
F = MDW
So, a spatial filter consists in building a filtering matrix
which is a projection matrix (ie. FF = F). The signal can
then be directly filtered by applying this matrix:
Projection matrices have interesting properties: they are
diagonalizable and the eigenvalues are all 1 or 0. They can
then be defined by two eigenspaces E1and E2of respective
dimensions n1and n2(n1+n2= n). The subspace, E1corre-
sponding to cerebral sources is defined by ∀x ∈ E1, Fx = x
and the subspace E2, corresponding to artifactual sources is
defined by ∀x ∈ E2, Fx = 0.
It is important to notice that the n1first line vectors of W
matrix (eq. 3) represent a base of the subspace E1 and the
n1 first column vectors of M are a base of the orthogonal
subspace of E2. In the same way, the n2last line vectors of
W matrix represent a base of the subspace E2and the n2last
column vectors of M are a base of the orthogonal subspace
of E1. The mixing of artifactual sources is then equivalent to
the separation of cerebral sources and reciprocally.
B. ICA automation
One of the major problems of ICA is the necessity to
manually identify each component as artifactual or not. Many
papers focus on this subject but most of them only focus on
ocular artifacts. Only the method of P. Le Van et al.  treats
all types of artifacts. It consists in extracting some features of
each source, characterizing location, frequencies, amplitude or
other properties. Then a Bayesian classifier is used to decide
from these features if a source is artifactual or cerebral.
IV. AFOP METHOD
The method proposed in this article is not based on ICA
but on Common Spatial Pattern (CSP) theory. It consists in
constructing a filtering matrix using a training step to learn
distribution of artifactual activity. On this purpose, two EEG
periods are compared. The first one contains resting EEG and
the second one contains EEG contaminated with artifacts. The
artifactual sources are then defined by the sources with the
most important increase of the variance between rest instant
and artifact instant. On the other hand, the EEG sources are
the sources with approximate constant variance.
This method has to learn the most common artifacts and
requires a protocol realized on each beginning of the recording.
During two minutes, the patients have to carry out four types of
artifact several times. This period is then compared to a resting
period (without artifacts) for the learning step. The four types
of artifact are: eye blinks, eye movements, jaw clenching and
B. Standard AFOP
The AFOP method  is split in two steps. The first
one consists in determining a subspace E1 by learning the
separation of cerebral sources with Common Spatial Pattern
(CSP) method. The second step consists in determining the E2
subspace with a least squared method on a sliding window.
This E2subspace corresponds to the best distribution of the
cerebral sources defined by E1.
For the first step, the CSP method find sources whom
variance increases the less between two learning periods corre-
sponding to matrix V1and V2. The two covariance matrices
are then calculated by C1 = V1VT
subspace E1 correspond to sources increasing the less and
it is defined by the eigenvectors ([w1,w2,w3,...] = WT)
corresponding to the smallest eigenvalues of:
The eigenvalues correspond to the variance ratios between
the two periods. It is possible to only select eigenvectors whom
eigenvalues are inferior to a threshold (empirically fixed to 2.5)
in order to determine automatically the dimensions n1and n2
of each subspace.
Once the subspace E1is defined, the orthogonal subspace
of E2is determined by finding the best distribution of cerebral
sources using a least squared method on the period to filter. A
sliding window of 20 seconds (of signal matrix Vt) is used to
cover the entire recordings. The best distribution is given by:
M = CtWT(WCtWT)−1
where Ct= VtVT
that if the period Vtcorresponds to the period V1or V2, the
subspace E2is given by the n2last eigenvectors of eq. 5.
For the artifact filtering application, one can suppose that
both cerebral and artifactual sources are static due to the
fact that the brain and the muscles are always at the same
place. However, it is possible that some sources would not be
activated during the resting period of the learning step. For
example the sources of paroxysms may not appear during this
period. By the way, there is only the artifactual distribution
space that will be considered as constant. The sec. III-A
shows that considering the artifact distribution to be constant
is equivalent to considering the cerebral separation to be
1and C2 = V2VT
tis the covariance matrix. It can be noticed
C. Heart beat artifacts
The AFOP method does not learn artifacts which are
constant during the entire recording. This is the case of heart
beat artifacts due to the fact that they are as important during
the rest period as during the artifact period. It is then possible
to use the method of J. Wang  by determining the time of
each heart beat by using Electrocardiogram (ECG). The mean
of signals taken one second after each beat is computed in
order to minimize the amplitude of the sources which are not
synchronized with ECG. A CSP is then carried out between
the entire period to filter and the mean signals. Sources with a
variance ratio closed to 1 are artifactual sources and are added
to artifact repartition subspace (orthogonal to E1).
Fig. 1.Illustration of stability improvement
D. Stabilized AFOP
In the case where the artifact distribution is slightly different
to the one of the learning period, some instability may occur.
This can be explained by the fact that a small artifact appears
on cerebral sources and the AFOP method tries to rebuild the
original signal by incorrectly amplifying those sources. Fig.
1.b illustrates the result of instability problem. A meaningless
signal appears on channel FP1 and FP2 because the method
tries to rebuild the artifact.
This instability corresponds to a small angle between the
two subspaces E1and E2. It is in general due to the fact that
the cerebral sources are negligible compared to the artifactual
ones. There are then a lot of degrees of freedom to rebuild the
artifacts. These degrees of freedom can then be canceled by
making a Principal Component Analysis (PCA) on the current
period. Then, a reduced space based on the first components
of PCA (those of bigger eigenvalues) is considered and the
subspace E1is projected on this base. The normal process is
then carried out using this reduced space.
The number of kept components with PCA decreases from
n until the angle between the two subspaces is superior to a
threshold (empirically fixed to 7◦) or until too much signal is
canceled with PCA. Fig. 1.c illustrates result of this method.
E. Band frequency decomposition
The described method allows to spatially characterize the
artifacts. However, artifacts can be defined as well by their
frequencies. This is the reason why each of these steps can be
carried out on frequency band decomposition. The signals in
each frequency band are treated independently and when all
frequency bands are spacially filtered, the final EEG signal will
be their sum. The selected frequency bands are the followings:
• The band (1-8 Hz) corresponding to the neurologic band
∆ (1-4 Hz) and θ (4-8 Hz). There is few of these
signals on normal EEG but their presence often reveals
abnormalities. All the eye and electrode artifacts are
contained on this band.
• The band (8-13 Hz) corresponding to neurologic band α.
Most of the normal EEG signals is in this band, generally
poor in artifacts.
• The band β (> 13 Hz). There are very few EEG β signals
that can be measured on the brain surface. This band
contains mainly muscular artifacts but during paroxysms,
the spikes are on this band. Taking into account its width,
a more precise segmentation is carried out with frequency
bands (13-20Hz), (20-30Hz) and (> 30Hz).
(a) Artifacted original signal
(b) Filtered signal with ICA
(c) Filtered signal with stabilized AFOP with frequency band decomposition
Fig. 2.Examples of filtering results with ICA and stabilized AFOP
For the band (1-8 Hz), only ocular artifacts and heart beats
artifacts are considered so the matrix V2is made only by ocular
artifacts period. On the other bands, only muscular artifacts are
considered and the stabilization is less important than in low
Six recordings of epileptic adult patients have been treated
with this method and examined by an expert. Two of these
recordings contain spike wave paroxysms, and three contain
slow waves activities. Fig. 2(c) illustrates some examples of
result that can be obtained with Stabilized AFOP and fig. 2(b)
shows results that can be obtained by manual ICA. The ICA
filtering has been made on 20 second periods EEG with JADE
algorithm. The automatic method using ICA gives less good
result than manual ICA since these methods can only add
classification error of components.
This figure shows that both methods filter most of artifacts
and remove only a small quantity of EEG signals. It can
be noticed that muscular artifacts are often less filtered with
ICA method and they are approximately equivalent concerning
ocular artifacts. This can be explained by the fact that the
frequency band decomposition treats each frequency indepen-
dently. In general, ICA trends to badly filter short duration
artifacts because there is not enough time to learn them. By
the other way, AFOP cannot filter electrode artifacts because
the spatial distribution is always different. These examples
show that both AFOP and manual ICA reduce the amplitude
of pathological elements of about 25%. Given that the AFOP
method continually applies the filter, this reduce is present
even if there is no artifact contrary to ICA. However the results
of AFOP are a few better if there is a mixing of EEG and
artifacts, because ICA may not manage to correctly separate
the sources. On all realized tests, the α rhythm is never reduced
whereas ICA can reduce it in presence of many artifacts. The
last part of this figure shows that AFOP method perfectly filter
heart beat artifacts whereas it is not always the case with ICA.
This method has been programmed in Matlab using
EEGLab toolbox . The computation of AFOP is instan-
taneous and only the filtering step for the frequency band
decomposition takes about 10 minutes on a Pentium 4 3GHz,
for a 3 hours recording. The matlab programs and a demon-
stration with video-EEG are available on the author’s website
The AFOP method filters automatically a large variety of
artifacts. However, some instability can occur when artifacts
are too important compared to the EEG signal or when
the training is not precise enough. This paper has presented
a method forincreasing this stability. Using the frequency
band decomposition, the results are often better than those
realizable with ICA. Only mechanical movement artifacts is
badly filtered with this method due to the fact these artifacts
are not static. The study on long duration recording shows that
it is possible to filter a large amount of artifacts while keeping
most of the EEG signal.
The authors gratefully acknowledge the medical personnel
of Hospital St Vincent and St Philbert of Lille. They are also
obliged for the financial support of FRM.
 S. Boudet, L. Peyrodie, P. Gallois, and C. Vasseur, “Filtering by Optimal
Projection and application to automatic artifact removal from EEG,”
Signal Process., vol. 87, no. 8, pp. 1978–1992, 2007.
 P. Le Van, E. Urrestarazu, and J. Gotman, “A system for automatic artifact
removal in ictal scalp EEG based on independent component analysis and
Bayesian classification,” Clinical Neurophysioly, vol. 117, no. 4, pp. 912–
 Y. Wang, “Reduction of cardiac artifacts in magnetoencephalogram,”
Proc. of the 12th International Conference on Biomagnetism, 2000.
 T. Jung, S. Makeig, C. Humphries, T. Lee, M. McKeown, V. Iragui,
and T. Sejnowski, “Removing electroencephalographic artifacts by blind
source separation,” Psychophysiology, vol. 37, pp. 163–178, 2000.
 P. Common, “Independent component analysis, a new concept,” Sig. Pro.,
vol. 36, pp. 287–314, 1994.
 A. Delorme and S. Makeig, “EEGLAB: an open source toolbox for
analysis of single-trial EEG dynamics including independent component
analysis,” J. Neuroscience Methods, vol. 134, no. 1, pp. 9–21, 2004.