Neurophysiologie Clinique/Clinical Neurophysiology (2016) 46, 287—305

Disponible en ligne sur

ScienceDirect
www.sciencedirect.com

COMPREHENSIVE REVIEW/REVUE GÉNÉRALE

Methods for artifact detection and removal

from scalp EEG: A review
Les méthodes de détection et de rejet d’artefact de l’EEG de
scalp : revue de littérature

Md Kafiul Islam a, Amir Rastegarnia b,∗, Zhi Yang a

a
Department of Electrical and Computer Engineering, National University of Singapore, Singapore
b
Department of Electrical Engineering, University of Malayer, Malayer, Iran

Received 26 February 2016; accepted 7 July 2016

Available online 15 October 2016

KEYWORDS Summary Electroencephalography (EEG) is the most popular brain activity recording tech-
Ambulatory EEG; nique used in wide range of applications. One of the commonly faced problems in EEG recordings
Artifact removal; is the presence of artifacts that come from sources other than brain and contaminate the
Brain-computer acquired signals significantly. Therefore, much research over the past 15 years has focused on
interface (BCI); identifying ways for handling such artifacts in the preprocessing stage. However, this is still an
Empirical mode active area of research as no single existing artifact detection/removal method is complete
decomposition (EMD); or universal. This article presents an extensive review of the existing state-of-the-art artifact
Independent detection and removal methods from scalp EEG for all potential EEG-based applications and
component analysis analyses the pros and cons of each method. First, a general overview of the different artifact
(ICA); types that are found in scalp EEG and their effect on particular applications are presented.
Scalp EEG; In addition, the methods are compared based on their ability to remove certain types of arti-
Wavelet transform facts and their suitability in relevant applications (only functional comparison is provided not
performance evaluation of methods). Finally, the future direction and expected challenges of
current research is discussed. Therefore, this review is expected to be helpful for interested
researchers who will develop and/or apply artifact handling algorithm/technique in future for
their applications as well as for those willing to improve the existing algorithms or propose a
new solution in this particular area of research.
© 2016 Elsevier Masson SAS. All rights reserved.

∗ Corresponding author.
E-mail address: [email protected] (A. Rastegarnia).
http://dx.doi.org/10.1016/j.neucli.2016.07.002
0987-7053/© 2016 Elsevier Masson SAS. All rights reserved.
288 M.K. Islam et al.

Résumé L’électroencéphalographie (EEG) est une technique d’exploration du cerveau très

MOTS CLÉS utilisée dans une large gamme d’applications. L’un des problèmes couramment rencontrés
Analyse en dans les enregistrements EEG est la présence d’artefacts qui viennent de sources autres que
composantes l’activité cérébrale et contaminent significativement les signaux acquis. En conséquence, de
indépendantes ; nombreux travaux de recherche ont été effectués depuis les années 2000 pour identifier les
EEG ambulatoire ; moyens d’éliminer ces artefacts dans une étape de prétraitement du signal. Ceci est toujours
EEG de scalp ; l’objet de recherches actives, car aucune méthode existante de détection et rejet d’artefacts
Interface n’est parfaite et n’a pu faire l’objet d’un consensus. Cet article présente une revue détaillée et
cerveau-machine ; un état de l’art concernant les méthodes de détection et rejet d’artefacts à partir des enreg-
Mode de istrements EEG de scalp pour toutes les applications potentielles basées sur l’EEG et analyse les
décomposition avantages et les inconvénients de chaque méthode. Tout d’abord, un aperçu général des dif-
empirique ; férents types d’artefacts qui peuvent s’observer dans l’EEG de scalp et leur impact en fonction
Rejet d’artefact ; d’applications particulières sont présentées. Puis, les méthodes sont comparées en fonction de
Transformation en leur capacité à éliminer certains types d’artefacts et de leur valeur dans les différentes appli-
ondelettes cations pertinentes (seule une comparaison « fonctionnelle » est présentée et non l’évaluation
de la performance de ces méthodes). Enfin, les orientations futures et les défis des recherches
actuelles sont discutées. Cette revue devrait être utile pour les chercheurs intéressés à dévelop-
per et/ou à appliquer des algorithmes ou techniques de manipulation d’artefacts EEG dans leurs
travaux futurs, ainsi que pour ceux qui souhaitent améliorer les algorithmes existants ou de
proposer de nouvelles solutions dans ce domaine de recherche spécifique.
© 2016 Elsevier Masson SAS. Tous droits réservés.

Introduction complete solution yet available for this particular problem.

More specifically, a careful review of the relevant artifact
Electroencephalography (EEG) is a non-invasive recording detection removal algorithms/methods reveals that there is
technique that measures the electrical activity of brain a gap between designed algorithm and its target application.
by placing electrodes on the scalp [65]. Due to its non- Most of the available techniques are not application-specific
invasiveness and cost-benefit ratio, EEG has been the most and therefore unnecessary computational burden arises.
preferred method of brain recording in clinical studies, lab Considering this issue, this paper aims to provide a com-
experiments, patient health monitoring [36], diagnosis and prehensive survey on the existing state-of-the-art artifact
many other applications. Unfortunately, EEG recordings are detection and removal methods from scalp EEG for all poten-
often contaminated by different forms of artifacts, such as tial EEG-based applications. It is worthy to note that this
artifacts due to electrode displacement, motion artifacts, research deals with artifacts and their handling methods
ocular artifacts and EMG artifacts from muscle activity. found only in scalp EEG recordings, not stimulation arti-
These offending artifacts not only misinterpret the underly- facts or artifacts found in simultaneous EEG-fMRI recordings.
ing neural information processing but may also themselves There are several useful algorithms proposed in the litera-
be difficult to identify. For example, during patient mon- ture to remove artifacts from such EEG-fMRI signals, such as
itoring in a critical care unit or during epilepsy seizure [2,3,25]. Interested researchers can take a look at these ref-
detection, artifacts may increase the chance of false alarms erences for more information. In addition, since currently
[26,84]. Another example is during brain-computer interface there is no universal standard quantitative metric avail-
(BCI) applications, where artifacts can modify or alter the able for performance evaluation of existing artifact removal
shape of a neurological event (e.g. event-related potential methods,1 this paper does not report such performance eval-
or ERP) that drives the BCI system and that eventually results uation, but rather provides only the functional comparison
in an unintentional control of the device [100]. The same between methods.
problem may occur during sleep study [82] and diagnosis To this end, first we briefly introduce typical artifact
of other neurological disorders such as Alzheimer’s disease types that are found in scalp EEG. Then, we provide a com-
(AD) [13], schizophrenia [95], etc. Therefore, artifact detec- parative analysis of the existing methods/algorithms with
tion and removal is one of the most important preprocessing their advantages, limitations and application-specific chal-
steps for neural information processing applications. lenges. Finally, the future direction is discussed to provide
The variety of artifacts and their overlap with signals of application-specific solutions with reasonable complexity,
interest in both spectral and temporal domains, even some- optimized performance and most importantly with feasible
times in the spatial domain, makes it difficult for simple
signal preprocessing technique to identify them from EEG.
Therefore, the use of simple filtering or amplitude thresh- 1 There are a couple of articles [39,52] that proposed to use sim-
olds to remove artifacts often results in poor performance ulated EEG data for performance evaluation of any artifact removal
both in terms of signal distortion and artifact removal. So far, method in a quantitative manner. Interested readers who wish to
a large number of methods/algorithms have been developed explore the quantitative performance evaluation technique of any
for artifact detection and removal from EEG signals. How- artifact removal method are requested to consult the mentioned
ever, as we will discuss in this paper, there is no universal articles for more details.
Scalp EEG artifacts 289

solutions. We believe that this review paper can help

Table 1 Description of notations.
researchers to choose the most suitable artifact handling
method for a particular EEG-based application. Moreover, Symbol Description
it would also be useful for those researchers interested
in designing and implementing new methods/algorithms to TVD Total variation de-noising
handle artifacts in a more efficient way, keeping in mind the EIH Energy interval histogram
particular application. EAS Ensemble average subtraction
A list of symbols and notations commonly used in this PWC-PSVM Probabilistic SVM with pairwise coupling
paper is shown in Table 1. APF Adaptive predictor filter
The rest of this paper is organized as follows. Section OPTIMI Online predictive tools for intervention in
‘‘EEG and artifact characterization’’ introduces typical EEG mental illness
and artifact characteristics. Section ‘‘Existing artifact hand- RBF-ANN Radial basic function based artificial neural
ling methods’’ briefly describes the mechanism of all the network
existing methods for artifact detection and removal. Section FORCe Fully online and automated artifact
‘‘Comparison between methods’’ provides a comparative removal for BCI
analysis between the methods and their suitability for differ- SFA Signal fraction analysis
ent applications. Section ‘‘Discussion’’ discusses the current GSVD Generalized singular value decomposition
status of artifact handling software plug-ins and also pro- EDS Exponentially damped sinusoidal
vides future directions of this research. Finally, section RMVB Robust minimum variance beamforming
‘‘Conclusions’’ gives concluding remarks. STF-TS Space-time-frequency time/segment
GMDH Group method of data handling
PNN Polynomial neural network
DTT Decision tree technique
ARX Auto-regressive exogenous
EEG and artifact characterization WNN Wavelet neural network
CSPA Component subspace projection algorithm
EEG characteristics SR Spectral ratio
FLN-RBFN Functional link neural network with
EEG is the recording of the electrical activities from sur- adaptive radial basis function networks
face/scalp of the brain and typically described in terms FLNN-ANFIS Functional link neural network with
of rhythms and transients. The rhythmic activity of EEG is adaptive neural fuzzy inference system
divided into bands of frequency. Although the common EEG MARA Multiple artifact rejection algorithm
rhythms are delta, theta, alpha and beta waves, however, FOOBI Fourth-order Tensor method
recently the gamma wave comes into EEG analysis in cer- UBSS Underdetermined blind source separation
tain cases. Moreover, mu wave is also considered as a variant TDSEP Temporal de-correlation source separation
because of lack of association with dysfunction or diseases. LAMIC Lagged auto-mutual information clustering
The corresponding frequency bands of these waves are given ERP Event-related potential
in Table 2. CNR Contrast-to-noise ratio
EEMD Ensemble empirical mode decomposition
MCCA Multi-set canonical correlation analysis
WPT Wavelet packet transform
Artifacts Local SSA Local singular spectrum analysis
MSSA Multivariate singular spectrum analysis
EEG recordings are often contaminated by different forms CC Correlation coefficient
of artifacts. The artifacts in EEG recording are of various RRMSE Relative root-mean-squared error
types that come from different sources. In broad sense, arti- LPM Linear programming machine
facts in EEG can be originated from internal and external JBSS Joint blind source separation
sources and contaminate the recordings in both temporal PSNR Peak signal-to-noise ratio
and spectral domains with wide frequency band. Internal EAWICA Enhanced automated wavelet-ICA
source of artifacts are due to physiological activities of SSA Stationary subspace analysis
the subject (e.g. ECG, EMG/muscle artifacts, EOG) and its CBSS Constrained BSS
movement. External source of artifacts are environmental MI Mutual information
interferences, recording equipment, electrode pop-up and FASTER Fully automated statistical thresholding for
cable movement. Also some artifacts may present in sev- EEG artifact rejection
eral neighboring channels (global) while some of them can OSET Open-source electrophysiological toolbox
be found only in single-channel (local). In addition, some AAR Automatic artifact removal
artifacts appear as regular periodic events such as ECG or ADJUST Automatic EEG artifact detector based on
pulse artifacts (regular/periodic) while some others may be the joint use of spatial and temporal
extremely irregular. An example of artifact-contamination features
is illustrated in Fig. 1. BCI Brain-computer-interface
A summary of different artifact types and their origins is
provided in Table 3.
290 M.K. Islam et al.

Figure 1 Left: a scalp EEG segment where all channels are more or less contaminated with muscle activity during the 10 seconds.
Right: the 10-second scalp EEG recordings with 21 channels from a long-term Epilepsy Monitoring Unit (OSG EEG recorders, Rumst,
Belgium). The seizure EEG was contaminated with muscle artifacts and eye blinks. Muscle artifacts can be observed between 0 sec
and 3.9 sec on channels F7, T3, T5, C3, and T1 and between 5 sec and 10 sec on channels F8, T4, F4, C4, and P4 [16].

Table 2 EEG rhythms with their corresponding frequency

Artifact avoidance
bands.
Artifact avoidance is a preventive and precautionary way
Rhythm or EEG signal Frequency to avoid or minimize artifacts by instructing the subject to
transient component band (Hz) remain still and try to avoid unnecessary blinks, eye/body
movements and so on. Also by proper grounding of the EEG
Rhythm Delta <4
recorder, one can reduce the supply mains interference.
Theta 4—8
Although artifact avoidance is not the best way to get rid of
Alpha 8—13
artifacts completely, minimizing artifacts can reduce both
Beta 14—30
the data loss and the computational complexity. However,
Gamma > 30
based on applications, sometimes this is a very unrealistic
Mu 7.5—12.5
solution; e.g. in an ambulatory EEG monitoring or brain-
Transient Seizure and 0.5—30
computer interface (BCI) application. Moreover, there are
inter-ictal
several limitations to employ such approach since some of
activities
the physiological artifacts (e.g. ECG) are involuntary and
therefore cannot be avoided. In addition, the subject can-
not limit eye blinking or movement for a long period of
time, especially if the subject is a child. Therefore, there
Existing artifact handling methods will always be some artifacts present in the recording and
those should be handled in the digital signal processing
In this section, we present the different artifact handling domain.
methods found from extensive literature review.

Table 3 Different types of artifacts and their origins.

Artifact types and sources

Physiological/internal Extra-physiological/external

Ocular Cardiac Muscle Others Instrumental Interference Movement

Eye blink ECG pulse Chewing Gloss kinetic Electrode Electrical Head
Eye movement Swallowing Skin Displacement Magnetic movement
Eye flatter Clenching Respiration and pop-up Sound Body
REM sleep Sniffing Cable Optical movement
Talking movement EM waves Limbs
Scalp Poor ground movement
contraction Tremor
Other
movements
Scalp EEG artifacts 291

Artifact detection

Identifying artifacts is the first and most important step

for handling artifacts. Often the artifacts overlap with
EEG signals in both spectral and temporal domains such
that it becomes difficult to use simple filtering or straight
forward signal processing technique. In many applica-
tions, it is required to identify or separate artifacts in
real-time, therefore knowing both the artifact and sig-
nal characteristics is really necessary in order to detect
them faster. Detection of artifacts may refer to detecting
a particular epoch or detecting an independent component
to be artifactual after performing independent compo-
nent analysis, ICA (detail about ICA is given later in this
section).
Whether it should be detected in time domain or fre-
quency domain or even in both by utilizing time-frequency
Figure 2 Machine learning classification for identifying arti-
analysis, this decision depends on the type of artifacts
factual epoch from clean EEG epoch.
and/or type of applications. Some of other factors for select-
ing a detection method include:

Artifact removal
• availability of a reference artifact source;
• the number of available recording channels;
Artifact removal involves canceling or correcting the arti-
• the need for removing the artifacts after detection stage.
facts without distorting the signal of interest. This is
primarily done in two ways: either by filtering and regres-
A few existing methods adopted the idea of machine sion or by separating/decomposing the EEG data into other
learning for artifact separation from useful EEG signal by domains.
training a classifier with (supervised) or without (unsuper-
vised) labeled training datasets. Once artifactual epochs
are identified by applying a machine learning algorithm, Regression
such epochs are either highlighted as artifact annotator to Regression analysis [43,101], using a multi-modal linear
the clinicians for helping in decision making (e.g. epileptic model between observed and a reference signal, is a
seizure detection) or can be rejected before examina- traditional way of identifying artifactual samples and con-
tion from clinician or before sending to automated signal sequently removing such sample that do not belong to the
processing system [70]. model. Observed artifact-contaminated EEG signal and an
Machine learning techniques are mainly two types: super- artifact reference signal are common methods for remov-
vised and unsupervised learning. Among supervised learning ing some physiological artifacts such as ocular and cardiac
algorithms, two most popular methods used for classifica- artifacts.
tion between artifact and brain signals are artificial neural However, such regression analysis often fails when there
network (ANN) [11,38,40,57,83] and support vector machine is no reference channel available. In addition, EEG signal
(SVM) [6,44,70,71,85,87]. Among unsupervised learning, k- being non-linear and non-stationary process, linear regres-
means clustering and outlier detection are most common sion is not the best choice for analysis in such applications.
in this particular area of research [70]. A basic approach Moreover, it can only be used to treat few particular types
to classify artifact from EEG by using the machine learning of artifact, not all types.
classifier is shown in Fig. 2.

Blind source separation

Artifactual segment rejection One of the most popular artifact detection/removal
methods is based on blind source separation (BSS)
One way to reduce the effects of artifacts is to reject/cancel [33,43,62,86,97], which aims to extract the individual
the epoch or segment of EEG data which is labeled as arti- unknown source signals from their mixtures and possibly
factual. The major drawback of this method is that it also to estimate the unknown mixing channels using only the
removes important EEG information, which results in the information within the mixtures observed at the output of
loss of data [52,66]. This was an early technique of handling each channel with no, or very limited, knowledge about
artifacts, but nowadays with the introduction of recent sig- the source signals and the mixing channel. Let denote by
nal processing techniques, the preference is for techniques X the observed signals in multi-channel recordings, which is
for artifact removal or correcting them instead of rejecting assumed to be linear mixture of the sources, S with additive
the data epoch. However, in certain applications, this tech- white noise vector N, i.e.
nique can still work reasonably well, e.g. offline analysis or
during training of any classifier. X = AS + N (1)
292 M.K. Islam et al.

the temporal correlation (i.e. auto-correlation). Then

CCA attempts to find an ordered set of components
from maximum auto-correlation to least auto-correlation.
The component with least auto-correlation corresponds
mostly to artifacts. The advantages of CCA over ICA are
being automatic and more computationally efficient;
Figure 3 Illustration of blind source separation technique. • MCA: morphological component analysis (MCA) decom-
poses the recorded signal into components that have
different morphological characteristics where each com-
Then, the objective is to find an estimate of the linear ponent is sparsely represented in an over-complete
mixture matrix A, denoted by W by an iterative process and dictionary [91]. It is only applicable to certain known arti-
obtain an estimate for the source signals as follows facts whose wave shape or morphology are known and
Ŝ = WX (2) stored in a database. The efficacy of this method greatly
depends on the available artifact-template database. In
A basic BSS technique is illustrated in Fig. 3. The main [106,107], MCA is used to remove ocular artifacts and
assumption with BSS is that the number of sources can be some of the muscle artifacts originating from jaw clench-
at most (or lower) equal to that of observed channels and ing, swallowing, and eyebrow rising.
the sources need to be independent (for ICA) or maximally
uncorrelated (for CCA) from each other:
Time-frequency representation
• ICA: independent component analysis (ICA) is based on Time-frequency analysis of non-stationary time series data
blind source separation (BSS) technique where it is is quite popular in biomedical signal processing, e.g. in EEG
assumed that the sources are linearly independent. The signal processing. The reason of using simultaneous time and
major problem with ICA-based artifact detection and frequency domain analysis is because of the non-stationary
removal is that, it is often not automatic. It requires properties of this type of signal. Therefore, any momentary
manual intervention to reject independent components change in frequency values for any signal components (e.g.
(ICs) with visually detected artifacts after decomposi- either artifact or seizure) [76,90] can be captured in a par-
tion. However, it (i.e. artifact detection and removal) ticular temporal window. In [67], a time-frequency analysis
can be made automatic by labeling the ICs through some of ocular artifacts (OAs) including blinks and saccades found
features that can quantify the possibility of being artifac- in EOG have been reported where the results reveal that
tual. Such procedure is performed by combining ICA with frequencies up to 181 Hz can be present in a subject’s EOG
another complementary method such as Wavelet Trans- for certain tasks. This finding suggests that if EOG is used
form or Empirical Mode Decomposition (EMD) or using for ocular artifact removal from EEG, then EOG should be
classifier like SVM or even with a help of reference chan- sampled at least 362 Hz to avoid aliasing.
nel [110]. However, even in such case, the artifactual ICs The common time-frequency representation is based on
may also contain some residual neural signals. Therefore, the short-time Fourier Transform (STFT). This method is not
during signal reconstruction after completely rejecting so effective as it has uniform time-frequency resolution at
that particular IC, it introduces distortion to the neu- all frequencies. For EEG, since the bandwidth is around
ral signal. Another problem is that it cannot operate on 0.5—120 Hz (although most of the time we are only inter-
single-channel data, since the number of recording chan- ested in < 30 Hz) and many of the artifacts (specially motion
nels must be at least equal to the number of independent and ocular artifacts) appear in the lower frequency region
sources. The computational complexity is another fac- (< 10 Hz), therefore, it is required to have high frequency
tor that limits the choice of ICA for artifact removal in resolution in lower frequency region which STFT cannot
applications that require online/real-time implementa- provide. To address this issue, a wavelet-based approach can
tion of the algorithm. Finally, the involvement of iterative be used as the wavelet transform, and provides proportional
process in computing ICA algorithm makes it difficult resolution in each frequency band suitable for EEG signals.
to perform robustly. E.g. ICA may be useful to remove
global artifacts such as ocular artifacts [11,27,31,43,46] Wavelet transform
or sometimes other physiological artifacts. There are few The wavelet transform is a time-scale representation
works reported the use of modified [23] or constrained method that decomposes signal f(t) into basis functions of
ICA [1,41,79,86] for automated and better performance time and scale which are dilated and translated versions of
in artifact detection and removal; a basis function (t) called mother wavelet [51]. Translation
• CCA: canonical correlation analysis or CCA is another BSS is accomplished by considering all possible integer transla-
method for separating a number of mixed or contaminated tions of (t) and dilation is obtained by multiplying t by
signals that uses second-order statistics (SOS) to generate a scaling factor, which is usually factors of 2. The follow-
components derived from their uncorrelated nature. By ing equation shows how wavelets are generated from the
looking for uncorrelated components, the approach uses mother wavelet:
a weaker condition than statistical independence sought
by the ICA algorithm. ICA does not take temporal cor- j,k (t) = 2j/2 (2j/2 t − k) (3)
relations into account while CCA addresses this point by
being capable of finding uncorrelated components [91]. So where j indicates the resolution level and k is the translation
the spatial correlation being zero while it optimizes only in time. This is called dyadic scaling, since the scaling factor
Scalp EEG artifacts 293

Table 4 Process flow of EMD algorithm to generate IMFS.

Input: data sequence s[n]

1. Identify all the local extrema

2. Separately connect all the maxima and minima with
natural cubic spline lines to form the upper, u[n], and
Figure 4 An example structure of 2-level decomposition by
lower, l[n], envelopes
discrete wavelet transform.
3. Find the mean of the envelopes as z[n] = [u[n] + 1[n]]/2
4. Take the difference between the data and the mean as the
is taken to be 2. Wavelet decomposition is a linear expansion proto-IMF, h[n] = s[n] — z[n]
and it is expressed as 5. Check the proto-IMF against the definition of IMF and the

+∞ 
+∞ 
+∞ stoppage criterion to determine if it is an IMF
f(t) = [ck (t − k)] + dj,k (2j t − k) (4) 6. If the proto-IMF does not satisfy the definition, repeat
k=−∞ k=−∞ j=0 step 1 to 5 on h[n] as many time as needed till it satisfies
the definition
where (t) is the scaling function and ck and dj,k are the
7. If the proto-IMF does satisfy the definition, assign the
coarse and detail level expansion coefficients, respectively.
proto-IMF as an IMF component, c[n]
A wide variety of functions could be chosen as the mother
8. Repeat the operation step 1 to 7 on the residue,
wavelet as long as following equation is satisfied:
q[n] = s[n] — c[n], as the data
 +∞
9. The operation ends when the residue contains no more
(t)dt = 0 (5) than one extremum
−∞

There are several techniques based on wavelet theory,

such as wavelet packets, wavelet approximation and decom- intrinsic mode functions (IMF) with well defined instanta-
position, discrete and continuous wavelet transform, and neous frequencies [58,94]. There are two basic conditions
so forth. The most commonly used technique is Discrete to be an IMF: (i) the number of extrema must be equal (or
Wavelet Transform (DWT). The DWT is derived from con- at most may differ by one) to the number of zero cross-
tinuous wavelet transform with discrete input. The relation ings (ii) any point, the mean value of the two envelopes
between input and output can be represented as defined by the local maxima and the local minima has to
be zero [58]. The general process flow of EMD algorithm

N

xa,L [n] = xa−1,L [2n − k]g[k] (6) is shown in Table 4. EEMD: it is an enhanced version of
k=1
EMD (enhanced empirical mode decomposition) and inspired
from the fact that EMD algorithm is very sensitive to noise,

N
which often leads to mode mixing complication. Therefore,
xa,H [n] = xa−1,L [2n − k]h[k] (7) EEMD is proposed which uses an average number of ensem-
k=1 bles (IMFs) from EMD as the optimal IMFs thus it provides a
where g[n] is a low pass filter just like scaling function and noise-assisted data analysis method [94].
h[n] is a high pass filter just like mother wavelet function.
Briefly, discrete wavelet transform is entering of a signal into Adaptive filtering
a low pass filter to get the low frequency component and An adaptive filter is a system with a linear filter that has
into a high pass filter to get the high frequency component. a transfer function controlled by variable parameters and
An example structure of 2-level decomposition by discrete a means to adjust those parameters according to an opti-
wavelet transform is shown in Fig. 4 [51]. mization algorithm [89]. The filter weights can adapt based
Once the signal is decomposed into detail and approxi- on the feedback from output of the system and it requires
mate coefficients, thresholding is applied on the coefficients a reference input to compare the desired output with the
to denoise the signal from artifacts. Then the new sets of observed output. An improved adaptive filtering by opti-
coefficients (all detail with final level approx. coefficients) mal projection which is based on common spatial pattern
are added up to reconstruct back the artifact-reduced sig- for artifact removal is mentioned in [9,10], especially for
nal. epilepsy patient’s EEG [74]. Let s[n] denote the observed
signal which is combination of the original EEG, x[n] and
Empirical mode decomposition additive artifact r[n]. Then, if the artifact source v[n] is
EMD is an empirical and data-driven method developed to available from a dedicated channel (e.g. EOG or ECG); an
perform on non-stationary, non-linear, stochastic processes adaptive algorithm (e.g. LMS, RLS, etc.) can be used to
and therefore it is ideally suitable for EEG signal analysis and derive an artifact-free EEG, x  [n] given that the desired EEG
processing. However, the computational complexity of EMD and artifact signal are independent (or at least uncorrelated
is quite heavy, so may not be suitable for online applications. [91]). An illustration of the use of adaptive filter for EOG
Moreover, the theory behind EMD is still not complete and artifact removal is shown in Fig. 5.
so far used in empirical studies, therefore it is difficult to
predict its robustness in all EEG recordings. Principal component analysis (PCA)
EMD algorithm decomposes a signal, s[n] into a sum PCA is a type of spatial filter that transforms the time
of the band-limited components/functions, c[n] called domain datasets into a different space by rotating axes in
294 M.K. Islam et al.

Figure 7 Process flow of the hybrid BSS-SVM algorithm.

signal. Hence, the multi-channel datasets are trans-

Figure 5 Typical use of adaptive filtering in canceling phys- formed into ICs or CCs and then possible artifactual
iological artifacts with available artifact source channel as component is decomposed by wavelet transform to dif-
reference. ferent frequency bands of detail coefficients. After that,
the artifactual coefficients are denoised by threshold-
an N-dimensional space (where n is the number of varia- ing, which eventually preserve the residual neural signals
bles or EEG channels) such that each dimension in the new of low amplitude after thresholding the higher artifac-
space has minimum variance and the axes are orthogonal to tual segments. The related articles are [14,34,50,52] for
each other [17]. PCA reduces data dimension and highlights wavelet-ICA, [109] for wavelet-CCA. On the other hand,
specific features of data, which is usually difficult to iden- there are similar hybrid methods that can be applied to
tify in the spatially unfiltered data as the new components single-channel EEG data by reversing the order of wavelet
are created by weighted combinations of all EEG channels. transform and BSS blocks. For example Calcagno et al. and
Two recent papers proposed artifact removal method based Mammone and Morabito [12,52] reported artifact removal
on PCA: Turnip [98] reported the use of robust PCA after by first decomposing signal into wavelet coefficients then
preprocessing is done based on wavelet de-noising and band- artifactual coefficients are passed through BSS block to
pass-filtering; while Turnip and Junaidi [99] compared PCA separate artifacts from neural signal. However, typically
with ICA for artifact removal and found ICA performs better the prior way is more known to the research community
than PCA. Both these papers have evaluated their method s wavelet enhanced ICA or wavelet enhanced CCA. An
qualitatively; therefore, it is not possible to comment exclu- example of such method is shown in Fig. 6. Please note
sively on the efficacy of PCA in detecting and removing that the type of wavelet transform can be DWT, CWT, SWT
artifacts. One important limitation of PCA (or SVD) is that or sometimes WPT [8];
it fails to separate/identify ocular or similar artifacts from • EMD-BSS: this hybrid method involves BSS with EMD
EEG when amplitudes are comparable since PCA depends on instead of wavelet transform. The difference is that usu-
the higher order statistical property [79]. ally the first stage is to decompose the signal into IMFs
by EMD or EEMD and then apply BSS (ICA or CCA) on the
Hybrid methods IMFs to identify artifactual component followed by reject-
In recent years, researchers have been keen to utilize the ing the artifactual IC or CC. The general process flow of
advantages of different methods by combining them into this hybrid method is also shown in the same Fig. 6. Such
a single method for artifact detection and removal, i.e. a methods are reported in [16,94,108];
hybrid method which has two or more stages. Some of these • BSS-SVM: Shoker et al. [87] reported a hybrid BSS-SVM
methods are discussed below: algorithm for eye blink and ECG artifact removal where
certain carefully chosen features are extracted from
• wavelet-BSS: this hybrid method formed by integrat- separated source components and then fed into a SVM
ing two popular methods: wavelet transform and blind classifier to identify artifact components followed by
source separation is mainly inspired from the fact that removing them. Finally, the rest of the source compo-
only BSS-based separation of artifactual components (e.g. nents are re-projected to reconstruct artifact-free EEG.
ICs) is often erroneous since the separated artifactual The whole system is illustrated in Fig. 7;
component also contains residual neural information. • REG-BSS: Klados et al. [43] reported a hybrid methodology
Therefore, completely rejecting such component will by combining BSS and regression-based adaptive filtering
introduce significant distortion in reconstructed EEG (with vEOG and hEOG as reference channels) for rejection

Wavelet-enhanced ICA/CCA (wICA or wCCA)

Multi-channel ICA Multi-channel

Denoising Inverse Inverse
Artifactual l or DWT Reconstructed l
(Thresholding) DWT CCA
Data CCA Data

EMD-ICA or EMD-CCA

Single-channel ICA Inverse Inverse

Single-channel
Artifactual l EMD or Thresholding ICA/CCA EMD
Reconstructed l
Data CCA Data

Figure 6 General process flow of EMD-BSS and wavelet-BSS methods.

Scalp EEG artifacts 295

Figure 8 Process flow of the hybrid REG-BSS methodology.

Removal performance

The performance evaluation of artifact removal methods

found in the literature is always problematic. It can be
done either by visually by expert(s) which is subjective (not
Figure 9 Process flow of the hybrid ICA-ARX methodology. standard) or by synthetic/semi-synthetic data (but uncer-
tainty of reconstructed data whether perfectly realistic or
not). Since there is neither any ground truth data available
of ocular artifacts as shown in Fig. 8. Similar techniques nor any universal or standard quantitative metric(s) used
have been used by [31] to remove ocular artifacts by com- in the literature that can capture both amount of artifact
bining ICA and adaptive filtering. Another hybrid approach removal and distortion. Therefore, it is quite difficult to
combining ICA and Auto-Regressive eXogenous (ARX) was compare different artifact removal methods based on their
proposed by Wang et al. [102] to remove ocular artifacts ability to remove artifacts since very few quantitative eval-
robustly as shown in Fig. 9. In this method, ARX is used uations have been reported in the literature. Most of the
to reduce the negative effect induced by ICA by building published articles evaluated their method in terms of some
the ARX multi-models based on the ICA correlated sig- qualitative plots. In addition, very few of them quantified
nals and the reference EEG that are selected prior to the the distortion to desired EEG signals due to the removal
artifact-contamination; effect. Therefore, it is not fair to tell which performs best
• other hybrid methods: Nguyen et al. [63] report EOG arti- based on the study.
fact removal using a hybrid method combined of Wavelet
decomposition and Artificial Neural Network and termed
as Wavelet Neural Network (WNN) where the reference Automatic or semi-automatic
EOG channel is only required during training of ANN clas-
sifier. A method combining DWT and ANC (Adaptive noise Most of the EEG-based applications require automated infor-
canceler) is proposed in [73] to remove ocular artifacts mation processing, particularly when it is an online/real-
where the OA reference is derived from DWT decomposi- time implementation. In addition, manual identification
tion and then used in the adaptive filter as reference. On of artifactual component or epoch is very time-consuming
the other hand, Navarro et al. [60] used the combination and laborious for multi-channel long-term data sequences.
of EMD and adaptive filter (with RLS algorithm) to remove Therefore, many signal processing techniques have been
ECG artifacts from EEG recordings. The authors in [38] proposed, and some useful a priori signal or artifact
presented a new way to remove EOG and EMG artifacts statistics/characteristics have been utilized. Among them,
from EEG by using a hybrid combination of functional link BSS-based techniques can sometimes be semi-automated
neural network (FLNN) and adaptive neural fuzzy infer- because of identification of artifactual component may
ence system (ANFIS). The ANFIS usually has two parts: require some training or parameter selection/tuning.
a non-linear antecedent and a linear consequent; how- Although there are few papers available that propose auto-
ever, in their proposed system, the second part is replaced mated identification of ICs after ICA [104,111]; however,
with the FLNN to enhance the non-linear approximation they both require training samples for supervised classifi-
ability. Then an adaptive filtering algorithm adjusts the cation and in addition requires an extra information in the
parameters of both ANFIS and FLNN. form of contact impedance measurement [31]. If the method
involves ICA for automatic detection of artifacts, then there
Statistical features has to be another stage (or method) in order to make the
Several statistical features [37,57,66] are used in machine whole process automated.
learning classifier or during threshold calculation in
wavelet/EMD/ICA-based methods for separating or identi-
fying artifacts from EEG signal of interest. Some of such Real-time/online implementation
features are discussed in Appendix A.
Online/real-time implementation requires the algorithm to
be fast enough and to have low-enough complexity for
Comparison between methods such application. Here, online implementation refers to
the algorithms implemented in software platform capable
In order to compare different artifact handling methods of online/real-time processing, not in hardware platform.
qualitatively, several factors need to be considered that can However, some EEG-based applications such as wireless
evaluate the pros and cons of these methods. Such factors ambulatory EEG monitoring may require on-chip implemen-
are described as follows: a detailed comparison between the tation of the artifact detection/removal algorithm. In such
existing artifact detection and removal methods in the liter- cases, the computational complexity has to be minimal,
ature found from recognized journals is provided in Table 5. which is a great challenge, and so far to the best of our
296 M.K. Islam et al.

Table 5 Comparative analysis of artifact removal methods found in literature published in recognized journals.

Articles Type of Method Online/ Automated Reference Multi/ Application

artifacts real-time single-
channel

Shoker Eye blink ECG BSS-SVM NIA Y N Multi General; e.g.

et al. [87] (SOBI-SVM) ERP analysis
Park et al. ECG EIH-EAS Y Y N Single General; e.g.
[72] sleep/wake
state or
epilepsy
monitoring
Hamaneh EKG ICA-CWT N/A Y Template Multi General; e.g.
et al. [34] epilepsy
monitoring
Shao et al. Eye ICA-weighted N/A Y Template Multi General
[85] blink + ECG PWC-PSVM
Zhao et al. Ocular DWT-APF Y Y N Single Monitor mental
[110] health (OPTIMI),
portable
applications
De Clercq Muscle CCA N N N Multi Epilepsy
et al. [20] monitoring;
applied on ictal
datasets
Ng et al. EOG + EMG SOBI-SWT N/A Y N Multi ␮ rhythm
[62] extraction
Mateo et al. Ocular RBF based N/A Y EOG channel Single General
[54] ANN (vEOG + hEOG)
Anderson EOG + 60-Hz GSVD-SFA May be N EOG channel Multi BCI; mental task
et al. [4] noise
Van Huffel Muscle + 50- SVD N/A N N Single/multi Ictal EEG
et al. [19] Hz
noise
Daly et al. Head ICA N Semi- Accelerometer Multi General; BCI
[18] movement automated
Noureddin EOG + Blink Adaptive Y Y Eye Tracker Multi General
et al. [68] Filter (RLS
and H␣ )
Peng et al. Ocular DWT-ANC May be Y N Single OPTIMI,
[73] portable
applications
Nazarpour Blink STF-TS-RMVB Y Y N Multi General
et al. [61]
James et al. Ocular cICA Y Y Derived Multi Seizure analysis
[41] reference
Schetinin ECG, EOG, PNN-GMDH- N/A Y Template Multi Sleeping
et al. [83] muscle, and DTT newborns
electrode
noise
Mahajan Eye blink ICA-DWT with N/A Y N Multi General
et al. [50] statistics
Kierkels EOG Kalman filter N/A Y Eye tracker Single General
et al. [42]
Sweeney Motion EEMD-CCA N/A Y N Single Ambulatory
et al. [94] single-channel
applications
Wang et al. Ocular ICA-ARX N/A Y N Multi General
[102]
Burger et al. EOG ICA-WNN N/A N N Multi General
[11]
Scalp EEG artifacts 297

Table 5 (Continued)

Articles Type of Method Online/ Automated Reference Multi/ Application

artifacts real-time single-
channel

Klados et al. Ocular REG-ICA N N EOG Multi General

[43]
O’Regan Head Feature fusion N/A Y Gyroscope Single Ambulatory
et al. [71] movement (69) to SVM EEG: seizure
monitoring + BCI
Ma et al. Ocular BSS-based N/A Y N Multi General
[49] CSPA
Ma et al. Muscle ICA-SR N/A Y N Multi General
[48]
Jafarifarmand Ocular Adaptive FLN- N/A Y ECG, EOG, Single/multi General
et al. [40] muscular RBFN-based EMG
and ECG filter (ANC)
Nguyen EOG WNN Y Y, training EOG only for Single Mental and
et al. [63] required training visual task
Hu et al. EOG and FLNN-ANFIS May be Y EOG, EMG Single/multi General
[38] EMG
Hartmann Most types Iterative N/A Y N Single/multi Epilepsy
et al. [35] Bayesian monitoring
Estimation
(MMSE)
Sameni Ocular Generalized N/A Y EOG Multi General
et al. [81] Eigenvalue
decomposi-
tion
Akhtar et al. Most types Spatially N/A Y May be Multi General
[1] cICA + Wavelet sometimes
de-noising
Molla et al. EOG Adaptive N/A Y Fractional Single General
[58] filtering Gaussian
(EMD-based noise
filter)
LeVan et al. Ocular, ICA + Bayesian N/A Y ECG Multi Ictal scalp EEG
[45] EMG, classification for epilepsy
movement diagnosis
Lawhern Ocular, AR model Yes Y N Single Real-time EEG
et al. [44] muscle, (fea- applications
movement ture) + SVM
Hallez et al. Muscle and BSS N/A Semi- N Multi Ictal EEG source
[33] ocular (CCA/spatial automated* imaging
cICA) + RAP-
MUSIC
Bhattacharyya All of them 26D N/A Y N Single Neonatal
et al. [6] features + bi- seizure
classification detection
Flexer et al. Ocular ICA N/A Semi- N Multi Blind subjects
[27] automated
Teixeira EOG + baseline Local N/A Y N Single General
et al. [96] drifts SSA + embedding
dimension
Ge et al. Ocular FOOBI based N/A Y N Multi Only for healthy
[28] on UBSS subjects; not
for epilepsy
298 M.K. Islam et al.

Table 5 (Continued)

Articles Type of artifacts Method Online/ Automated Reference Multi/ Application

real-time single-
channel

Nicolaou EOG, EMG and ECG TDSEP + LAMIC N/A Y EOG Multi Discovery and
et al. [64] analysis of ERP
Rashed-Al- Ocular EMD N/A Y Simulated Multi BCI
Mahfuz
et al. [77]
Guerrero- Ocular Adaptive N/A Y Fpl, Fp2, F7 Multi General
Mosquera filtering + ICA and F8
et al. [31] Electrodes
Mammone Ocular + muscle + EAWICA N Y N Multi General
et al. [52] electrical shift (wICA)
Winkler EOG + EMG TDSEP (based Y Y N Multi BCI
et al. on ICA) + LPM
[104]
Chen et al. Muscle EEMD-JBSS N/A Y N Single General + ictal
[16] EEG
Zeng et al. EOG SSA N N N Multi Diagnosis
[108] (BSS) + EMD

knowledge, no real-time hardware implementation has been ocular artifacts for portable EEG applications which is
performed. found to achieve lower MSE and higher correlation between
cleaned and original EEG in comparison with existing
methods such as wavelet packet transform (WPT) and
Single or multi-channel independent component analysis (ICA), discrete wavelet
transform (DWT) and adaptive noise cancellation (ANC).
BSS-based methods require multi-channels to function, the Another article [43] reported an automated ocular artifact
more number of channels the better for separating indi- removal method using adaptive filtering and ICA with the
vidual sources. Therefore, such methods cannot be used help of vertical (vEOG) and horizontal (hEOG) EOG channel
in low-channel (e.g. 4—6) or single-channel based appli- as reference. On the other hand, Flexer et al. [27] pro-
cations (e.g. in ambulatory monitoring of epilepsy patient posed an ICA-based ocular artifact removal method from
or ambulatory BCI-prosthesis). On the other hand, Wavelet blind subjects’ EEG utilizing both vertical and horizontal
transform and EMD-based techniques can work with single- EOG references.
channel analysis by decomposing a single data sequence into
multiple components (approx./detail coefficient for wavelet
decomposition and IMF for EMD). ECG
Authors in [21] proposed removal/reduction of ECG/cardiac
artifacts from EEG using a separate ECG reference channel.
Reference channel In [31], an automatic method based on a modified ICA algo-
rithm has been proposed that works for a single-channel EEG
Most of the available methods require a dedicated arti- and the ECG (as reference) which gives promising results
fact channel to be functional. In order to remove ocular or when compared with two popular methods that use a refer-
cardiac artifacts, the reference channel often provides sat- ence channel namely ensemble average subtraction (EAS)
isfactory complementary information to identify ECG/EOG and adaptive filtering. The other two articles proposed
artifacts. Besides, real-time contact impedance measure- their methods for application in neonatal EEG monitoring.
ment can provide the complementary information about Another paper [60] proposed a combination of EMD and
artifacts due to electrode pop, movement or loose con- adaptive filtering based method for ECG artifact removal
nection. Some movement tracking devices such as motion in preterm EEG and reported up to 17% improvement in cor-
captured camera, accelerometer and/or gyroscope can help relation coefficient between original and cleaned datasets
to detect motion artifacts. compared with removal by only adaptive filtering.

EOG Eye tracker

Many articles reported to remove EOG artifacts by the use Both Kierkels et al. [42] and Noureddin et al. [68] reported
of EOG reference channel [27,43,110]. In [110], a hybrid de- techniques for removal of ocular artifacts by using an eye
noising method has been reported that combines discrete tracker as reference. The advantage of using eye tracker
wavelet transformation (DWT) and an adaptive predictor is that it can reduce the undesired EEG distortion pro-
filter (APF) for automatic identification and removal of duced by using an EOG channel as reference since EOG
Scalp EEG artifacts 299

not only captures ocular events but also some frontal EEG from a high-density EEG recordings (248-channel) with the
events. Besides, in practical daily applications, the use of help of kinematics and kinetics information acquired from a
eye tracker removes the requirement of EOG electrodes 8-camera, 120 frames/s, motion capture system. The sub-
attached to the face. Results in [42] show significantly ject was asked to walk and run on a custom built, dual-belt,
improved performance in removing of only eye movement force measuring treadmill with two 24-inwide belts mounted
artifacts by combining Kalman filter with the eye tracker flush with the floor while simultaneously both brain and
information compared with three other popular methods body dynamics were recorded. The findings conclude that
namely Regression, PCA, and SOBI. On the other hand, high-density EEG is possible to use in order to study brain
Noureddin et al. [68] introduced an online algorithm for ocu- dynamics during whole body movements; and the artifact
lar artifacts (both movements and blink) removal from EEG from rhythmic gait events can be reduced by template
by utilizing a high-speed eye tracker (> 400 Hz) along with regression procedure.
the frontal EEG as reference instead of EOG channel. The
article used two adaptive filters (RLS and H) to prove the
Robustness
efficacy of their proposed technique, which was shown to
outperform the techniques using only EOG as reference.
Robustness is an important issue in developing any arti-
fact removal algorithm as artifacts are of diverse types and
Accelerometer contaminate the EEG differently in different recording envi-
There are few articles reported to have used accelerometer ronments. Some of the factors that should be considered for
recordings in conjunction with EEG recordings for detecting robustness include artifact-SNR, type of artifact, duration
motion artifacts [82,93]. In [82], it has been shown that of artifacts, subject-variability, environmental variability,
movement artifacts can be detected automatically using application-specificity.
an accelerometer with a developed algorithm based on AR
modeling and thus can increase the speed efficiency for
automatic computation of EEG model parameters compared Discussion
with manual detection of movement artifacts. Sweeney
reported in [93] that the use of accelerometer as refer- Current status
ence channel not only can detect motion artifacts but also
can remove them with the use of different filtering tech- Although significant amount of efforts has been made to
niques such as adaptive filters, Kalman filtering and Wiener develop methods for artifact detection and removal in EEG
filtering. applications, it is still an active area of research. Most of
them handle single type of artifact, many of them cannot
Gyroscope work for single-channel EEG, some of them require training
Authors in [71] proposed to detect different head movement data, some require a dedicated reference channel, some are
artifacts automatically by using a gyroscope as complemen- designed for general purpose applications that often leads
tary features in fusion with EEG features and finally with to overcorrection of data and some of them are not fully
the help of SVM, to classify artifacts from neural informa- automated. Some of the currently available major software
tion. The method is inspired by the realization of an artifact plug-in GUIs are discussed in Appendix B.
detection system for implementing with the point-of-care
REACT (Real-time EEG Analysis for event detection) tech- Future direction
nology that has potential application in the detection of
neurological events (e.g. seizure events) in adults. The arti- Here we present the future direction for handling artifacts
facts were generated for 10 different types of head-related by raising realistic issues, proposing some ideas and provid-
movements using 14-channel Emotiv EEG headset and the ing recommendation based on review of existing solutions.
movement time was recorded for validation during arti-
fact detection. The reported accuracy in terms of Avg. ROC
Probability mapping
areas was 0.802 and 0.907 for participant independent and
From the above literature review of existing solutions for
dependent systems respectively.
artifact handling, it is obvious that artifacts are of dif-
ferent types and not all types will play major role in all
Contact impedance measurement EEG-based applications. Sometimes, clinicians prefer man-
Bertrand et al. and Mihajlovic et al. [5,55,56] reported that ual event detection than automated algorithm for certain
by measuring the change in contact impedance due to head disease diagnosis (e.g. seizure detection). However, such
movements can help to estimate the motion artifacts and by manual analysis is also time-consuming. In such cases, if we
utilizing this information with an adaptive filter in combina- can give the users an option to choose which particular arti-
tion with band-pass filtering, the artifacts can be reduced facts they want to be detected and/or removed with what
significantly in real-time. The article also studies the effect amount (%) for each epoch or data-segment of duration 1-
of head movement artifacts on EEG recordings results in sec (depends on application), then the process would still
contaminating the spectral domain in < 20 Hz frequency. be automated with tuning facilities for the users either to
turn-ON or remain OFF if not required. In order to imple-
Motion captured camera ment such facility, a probability mapping of artifacts can
Authors in [32] proposed a channel and IC-based method be proposed (something similar to the idea of [105]) for
to remove movement artifacts during walking and running each epoch of data based on some statistical features to
300 M.K. Islam et al.

quantify the probability of an epoch to be artifactual. Then movement in an ambulatory environment [15]. Thus, it will
the user can opt for some threshold of probability above be easier to label both ground truth EEG and artifacts.
which he/she may want to remove artifacts while below
the threshold, to preserve the epoch as it is. Thus it is pos- Recommendation
sible to design automated artifact detection and removal In order to choose the right artifact handling method, we
algorithm, which is application-specific with tuning facil- need to consider the particular application, required spec-
ity for user. This would greatly enhance the signal analysis ification to be satisfied given the computational resources
process by avoiding the chance of removing important sig- and recording environment available. There are EEG appli-
nal information. In addition, it will reduce the unnecessary cations where only one or two types of artifacts affect the
computational resources and time by focusing on the desired later stage information decoding or processing, thus it is not
artifacts for detection/removal (i.e. only those types to be wise to attempt to identify and remove all the artifacts as
expected to affect the signal quality) and ignoring the rest other artifacts may not (or minimally) harm a particular sig-
of them. nal processing purpose. If any reference channel is available
in the targeted application, then regression or adaptive fil-
tering technique may be a preferred solution. In the case
Standard performance evaluation of ambulatory EEG monitoring, when number of channels
One of the important issues in evaluating the perfor- are fewer, no reference channel is available and wireless
mance of any artifact detection or removal method is EEG transfer preferred, in such case it is recommended to
that there is no universal standard quantitative metric use computationally cheaper method that can work without
for the researchers to use. Most of the methods men- reference and on single or few channels, e.g. wavelet-based
tioned in the literature use some qualitative time/frequency methods since BSS-based methods may not perform satisfac-
domain plot to evaluate the artifact removal performance tory with less number of channels. In some applications, if
or evaluated by the clinical expert. Sweeney et al. [92] it is possible to have some a priori knowledge about arti-
proposed a recording methodology for accurate evalua- facts and some training data available, and the application
tion and comparison between different artifact removal only require to identify artifacts not to remove them, then
techniques/algorithms which presented the EEG recordings machine learning based classifiers can be good choice. If the
of two separate but highly-correlated channels that allow EEG recording involves high-density channels, then PCA may
recording both artifact-contaminated and artifact-free sig- be preferred to reduce the dimensionality before applying
nal simultaneously. It also presented a tagging algorithm any artifact removal methods, such as BSS-based methods.
employing two accelerometers for generating a quality-of- If the application is based on offline analysis, then we can
signal (QOS) metric, which can be used to for multiple afford some computational expensive techniques such as ICA
purposes such as classification of motion artifacts, activation or EMD.
of artifact removal technique only when required and iden-
tification of the artifact-contaminated epochs. Thus, this
approach can provide accurate measurements of quantita- Conclusions
tive metrics for fair performance evaluation.
However, such methodology still requires intervention to An extensive analysis of the existing methods for arti-
the recording technique and also extra reference channel fact detection and removal has been presented with their
for accelerometer data, which may not be feasible in every comparison, advantages and limitations. The research on
application (e.g. portable EEG recordings). Although it is handling artifacts present in the typical EEG recordings is
highly encouraged for the removal performance to be evalu- still an active area of research and none of the existing
ated by the domain experts, however, such evaluation varies methods can be considered as the perfect solution. Most
from one expert to another and still is manual and/or qual- of the solutions do not consider the particular application,
itative evaluation. Therefore, it is an urge to have a single therefore, not optimized for that application. Although,
standard evaluation method consists of both qualitative and most of the removal algorithms provide good performance,
more importantly quantitative metrics or ways for evaluat- however, they are only suitable for offline analysis because
ing the performance in a more realistic and fair manner. of their high computational complexity and unsupervised
nature. Some of them even require a dedicated reference
channel, which is not feasible for some applications. Fur-
Ground truth data ther studies are required to characterize the properties of
Another reason of not being able to evaluate artifact commonly encountered artifacts and to observe the effects
removal performance fairly is that the lack of availability of their contamination to the desired later stage signal
of ground truth data. It’s now equally important to have processing/analysis. Some applications may only require to
a public database with sufficiently long-term EEG recor- identify artifacts and not to remove them, e.g. in appli-
dings without or minimal artifacts to be used as a ground cations where classification/identification of two classes
truth data. Besides such, an acceptable mathematical model are required. In such cases, a more realistic mathematical
to generate basic EEG rhythms and finally integrate them model of the desired event(s) to be identified is essential in
to simulate an EEG sequence with standard 10—20 system order to easily ignore other non-brain signals (i.e. artifacts
EEG channels is required for quantitative evaluation of any or interferences). Finally, the future direction will be to
existing/future artifact removal methods. In addition, more provide application-specific solutions with reasonable com-
study is necessary to characterize as much as possible of all plexity, optimized performance and most importantly with
artifact types, specially the motion artifacts for different feasible solutions.
Scalp EEG artifacts 301

Disclosure of interest in 0.5—29 Hz [26]. These bands and their FFT or spectral
power are useful features for separating artifacts from EEG.
The authors declare that they have no competing interest. FFT, F: Fast Fourier Transform or FFT is the frequency
representation of time domain signal values. For feature
extraction, we have used the mean of the absolute of FFT
Acknowledgments values for each epoch computed over the entire frequency
range of EEG signal (i.e. 0—128 Hz).
This work was supported by A*STAR PSF Grant R-263-000-
699-305 and NUS YIA Grant R-263-000-A29-133. F = mean(abs[FFT (k)]) (12)
Maximum FFT, Fmax : This feature is the maximum or peak
Appendix A. Statistical features value of the absolute of FFT values.
Fmax = max(abs[FFT (k)]) (13)
Time Domain Features
Spatial Features
Entropy, H: is a measure of uncertainty of information
content [78], of a discrete random variable x with possible
values x1 , ..., xn , can be calculated as: Spatial distribution or topographic mapping helps to iden-
tify the origin of many artifacts (e.g. ocular artifacts are
H(x) = E[− ln(P(x))] (8) dominant in frontal EEG channels) [93]. In addition, some
artifacts may appear in several nearby channels (global arti-
Here E is the expected value operator and P(x) is the facts such as eye blink) where some appear only in one
probability mass function of x. channel (i.e. local artifacts). Therefore, spatial features
Kurtosis, Kr: Kurtosis is the measure of ‘‘peakedness’’ of along with their spectral content are important to identify
probability distribution function [50] and is calculated for a artifacts from EEG signals [57,88].
real-valued random variable x as follows
4 Appendix B. Software plug-ins
Kr[x] = (9)
4
where  and  are the mean and standard deviation of ran- FORCe
dom variable x.
Line Length, L[n]: Line length, a signal feature for Fully Online and automated artifact Removal for brain-
seizure onset detection as reported by [24,59], for a discrete Computer interfacing or FORCe is the most recent method
time signal x[k] can be represented by, reported in [18] that is based on a unique combination of
WT, ICA and thresholding. Compared with two other state-

n
of-the-art methods namely LAMIC and FASTER, FORCe has
L[n] = abs[x[k − 1] − x[k]] (10)
been shown to outperform them significantly and is capa-
k=n−N
ble of removing different types of artifacts including eye
where N is the time window length. Here N = 1 sec. blink, EOG and EMG. One of salient features of FORCe is
Maximum, M: It is the maximum or peak value of an that it doesn’t require any reference channel and can oper-
epoch and noted down as a feature. ate on fewer numbers of channels which makes it suitable
NEO,  : The ability of Non-linear Energy Operator (NEO) for ambulatory EEG applications.
to enhance signal’s transition or large amplitude event
[53,57,75] is sometimes considered as feature for seizure FASTER
classification. The NEO operator  applied to a discrete time
variable x[n] is calculated as follows
FASTER stands for Fully Automated Statistical Thresholding
2
 [x[n]] = x[n] − x[n + 1]x[n − 1] (11) for EEG artifact Rejection which is an unsupervised algo-
rithm for parameter estimation in both EEG time series and
Usually the mean and/or variance of  [x[n]] for each in the ICs of EEG [66]. The achieved sensitivity and speci-
epoch are used as feature(s). ficity is > 90% for detection of EOG and EMG artifacts, linear
trends and white noise in the contaminated channels.
Frequency Domain Features
LAMIC
Spectral features along with temporal or spatial features are
often used for EEG classification. As mentioned before, EEG Lagged auto-mutual information clustering (LAMIC) is a clus-
rhythms have different spectral bands, therefore sometimes tering algorithm developed for automatic artifact removal
the relative power in those bands are used as features for from EEG [64]. The method involves data decomposition
classifier training. It is important to note that apart from the by a BSS algorithm called TDSEP (Temporal De-correlation
rhythms, there are recently reported High Frequency Oscil- source SEParation), which is a temporal extension of ICA.
lations (HFO having band of 80—200 Hz), Ripple (200—600 Hz) Then the components are clustered using the similarity of
bands present in EEG. In addition, the frequency band of typ- their lagged Auto-Mutual Information (AMI). This is inspired
ical Scalp EEG is 0.05—128 Hz while epileptic seizure appears from the fact that EEG and artifacts are different from their
302 M.K. Islam et al.

temporal dynamics point of view. The clustering procedure artifact-specific spatial and temporal features to automat-
follows the usual steps of hierarchical clustering. ically identify the artifactual ICs after ICA is performed.
Four different artifact types (i.e. eye blink, vertical eye
PureEEG movement, horizontal eye movement and generic disconti-
nuities) are chosen for extracting features such as temporal
kurtosis, spatial average and variance difference, maximum
This is an automatic EEG artifact removal algorithm for
epoch variance, spatial eye difference. The key feature of
epilepsy monitoring that based on a neurophysiological
ADJUST is that it is entirely automated and unsupervised
model by utilizing an iterative Bayesian estimation scheme
with reported accuracy of 95.2% in classifying all of the four
[35]. The method targets to remove most of the artifact
artifacts. It can also successfully reconstruct the clean ERP
types and does not require any manual intervention. The
topographies from heavy artifact-contamination.
authors reported the performance of PureEEG from two
independent clinical experts perspective and its found to
be significantly improving the readability of EEG recordings
after artifact removal. PREP Pipeline

OSET The PREP pipeline is a standardized preprocessing tool for

large-scale EEG analysis [7], which includes an automatically
OSET is an Open-Source Electrophysiological Toolbox for generated report for each dataset processed. The salient
biomedical signal generation, modeling, processing, and features of this toolbox include (i) removal of line-noise
filtering [80]. It can remove cardiac artifacts from any without incorporating typical filtering technique, (ii) ref-
bioelectrical signal including EEG. It can also handle and erencing the signal robustly, and (iii) identification of bad
remove EOG artifacts from multi-channel EEG using tech- channels relative to the reference.
niques based on semi-blind source separation.

MARA Makoto’s Preprocessing Pipeline

Multiple Artifact Rejection Algorithm (MARA) is an open- This pipeline is Makoto Miyaksohi’s personally recom-
source MATLAB-based EEGLAB2 plug-in which automatically mended EEG preprocess pipeline [30], which is a forever
identify the artifact-contaminated independent compo- beta version. Interested readers are requested to con-
nents for artifact rejection [103,104]. The main part of sult the following link for more details: [https://sccn.ucsd.
MARA is a supervised machine learning algorithm that learns edu/wiki/Makoto’s preprocessing pipeline].
from labeled components by experts and utilizes six fea-
tures based on spatial, spectral and temporal domain. It can
handle any type of artifact.
FieldTrip
AAR
This is an open-source MATLAB toolbox for MEG and EEG
Automatic Artifact Removal (AAR), a MATLAB toolbox which analysis which offers advanced analysis methods of MEG,
can be integrated as a plug-in into EEGLAB, includes dif- EEG, and invasive electrophysiological data, such as time-
ferent artifact removal methods for removing only EOG frequency analysis, source reconstruction using dipoles,
and EMG artifacts [29]. In order to remove only EOG arti- distributed sources and beamformers and non-parametric
facts, regression-based methods such as least mean squares statistical testing [69].
(LMS), conventional re-cursive least squares (CRLS), sta-
ble re-cursive least squares (SRLS) and algorithms based
on the H norm are used. For removing both EOG and
EMG artifacts, spatial filters based techniques have been ERPLAB
adopted.
ERPLAB is also EEGLAB-compatible open-source toolbox for
analyzing ERP data, which has artifact rejection capability
ADJUST in both manual and automated manner [47].

ADJUST, reported by Mognon et al. [57], is an EEGLab sup-

ported plug-in for automated EEG artifact detection. This
algorithm is based on the combined use of stereotyped References

[1] Akhtar MT, Mitsuhashi W, James CJ. Employing spatially con-

2 EEGLAB is an open-source MATLAB-based interactive GUI toolbox strained ICA and wavelet denoising, for automatic removal
for analyzing and processing continuous and event-related EEG, MEG of artifacts from multichannel EEG data. Signal Process
and other electrophysiological signals. It uses ICA, time-frequency 2012;92:401—16.
analysis, artifact rejection, event-related statistics and different [2] Allen PJ, Polizzi G, Krakow K, Fish DR, Lemieux L. Identi-
modes for visualizing the averaged or single-trial EEG data [22]. fication of EEG events in the MR scanner: the problem of
Scalp EEG artifacts 303

pulse artifact and a method for its subtraction. Neuroimage preprocessing refines neonatal seizure detection. Clin Neu-
1998;8:229—39. rophysiol 2011;122:2345—54.
[3] Allen PJ, Josephs O, Turner R. A method for removing imaging [22] Delorme A, Makeig S. EEGLAB: an open source toolbox for
artifact from continuous EEG recorded during functional MRI. analysis of single-trial EEG dynamics including independent
Neuroimage 2000;12:230—9. component analysis. J Neurosci Methods 2004;134:9—21.
[4] Anderson CW, Knight JN, O’Connor T, Kirby MJ, Sokolov A. [23] Devuyst S, Dutoit T, Stenuit P, Kerkhofs M, Stanus E. Cancelling
Geometric subspace methods and time-delay embedding for ECG artifacts in EEG using a modified independent compo-
EEG artifact removal and classification. IEEE Trans Neural Syst nent analysis approach. EURASIP J Advances Signal Process
Rehabil Eng 2006;14:142—6. 2008;2008:180.
[5] Bertrand A, Mihajlovic V, Grundlehner B, Hoof CV, Moonen M. [24] Esteller R, Echauz J, Tcheng T, Litt B, Pless B. Line length: an
Motion artifact reduction in EEG recordings using multichan- efficient feature for seizure onset detection. In: Proceedings
nel contact impedance measurements. In: Proceedings of IEEE of the 23rd Annual International Conference of the IEEE Engi-
Biomedical Circuits and Systems Conference (BioCAS). 2013. neering in Medicine and Biology Society. 2001. p. 1707—10.
p. 258—61. [25] Ferdowsi S, Sanei S, Abolghasemi V, Nottage J, O’Daly
[6] Bhattacharyya S, Biswas A, Mukherjee J, Majumdar AK, O. Removing ballistocardiogram artifact from EEG using
Majumdar B, Mukherjee S, et al. Detection of artifacts short and long-term linear predictor. IEEE Trans Biomed Eng
from high energy bursts in neonatal EEG. Comput Biol Med 2013;60:1900—11.
2013;43:1804—14. [26] Fisher RS, Vickrey BG, Gibson P, Hermann B, Penovich
[7] Bigdely-Shamlo N, Mullen T, Kothe C, Su KM, Robbins KA. The P, Scherer A, et al. The impact of epilepsy from the
prep pipeline: standardized preprocessing for large-scale EEG patient’s perspective I. Descriptions and subjective percep-
analysis. Front Neuroinformatics 2015;9:1—20. tions. Epilepsy Res 2000;41:39—51.
[8] Bono V, Jamal W, Das S, Maharatna K. Artifact reduction in [27] Flexer A, Bauer H, Pripfl J, Dorffner G, Using ICA. for removal
multichannel pervasive EEG using hybrid WPT-ICA and WPT- of ocular artifacts in EEG recorded from blind subjects. Neural
EMD signal decomposition techniques. In: Proceedings of IEEE Netw 2005;18:998—1005.
Acoustics, Speech and Signal Processing (ICASSP). 2014. p. [28] Ge S, Han M, Hong X. A fully automatic ocular artifact removal
5864—8. from EEG based on fourth-order tensor method. Biomed Eng
[9] Boudet S, Peyrodie L, Gallois P, Vasseur C. Filtering by optimal Lett 2014;4:55—63.
projection and application to automatic artifact removal from [29] Gomez-Herrero G. Automatic artifact removal (AAR) toolbox
EEG. Signal Process 2007;87:1978—92. v1. 3 (release 09. 12. 2007) for MATLAB. Tampere University
[10] Boudet S, Peyrodie L, Forzy G, Pinti A, Toumi H, Gallois P. of Technology; 2007.
Improvements of adaptive filtering by optimal projection to [30] Groppe DM, Makeig S, Kutas M. Identifying reliable inde-
filter different artifact types on long duration EEG recordings. pendent components via split-half comparisons. Neuroimage
Comput Methods Programs Biomed 2012;108:234—49. 2009;45:1199—211.
[11] Burger C, Heever DJVD. Removal of EOG artefacts by com- [31] Guerrero-Mosquera C, Navia-Vazquez A. Automatic removal
bining wavelet neural network and independent component of ocular artifacts using adaptive filtering and independent
analysis. Biomed Signal Process Control 2015;15:67—79. component analysis for electroencephalogram data. IET Sig-
[12] Calcagno S, Foresta FL, Versaci M. Independent component nal Process 2012;6:99—106.
analysis and discrete wavelet transform for artifact removal [32] Gwin JT, Gramann K, Makeig S, Ferris DP. Removal of move-
in biomedical signal processing. Am J Appl Sci 2014;11:57—68. ment artifact from high-density EEG recorded during walking
[13] Cassani R, Falk TH, Fraga FJ, Kanda PA, Anghinah R. and running. J Neurophysiol 2010;103:3526—34.
The effects of automated artifact removal algorithms on [33] Hallez H, De Vos M, Vanrumste B, Van Hese P, Assecondi S, Van
electroencephalography-based Alzheimer’s disease diagnosis. Laere KP, et al. Removing muscle and eye artifacts using blind
Front Aging Neurosci 2014;6:55. source separation techniques in ictal EEG source imaging. Clin
[14] Castellanos NP, Makarov VA. Recovering EEG brain signals: Neurophysiol 2009;120:1262—72.
artifact suppression with wavelet enhanced independent [34] Hamaneh MB, Chitravas N, Kaiboriboon K, Lhatoo SD,
component analysis. J Neurosci Methods 2006;158:300—12. Loparo K, et al. Automated removal of EKG artifact from
[15] Chang BS, Schachter SC, Schomer DL. Atlas of ambulatory EEG data using independent component analysis and con-
EEG. Academic Press; 2005. p. 56—74. tinuous wavelet transformation. IEEE Trans Biomed Eng
[16] Chen X, He C, Peng H. Removal of muscle artifacts from 2014;61:1634—41.
single-channel EEG based on ensemble empirical mode [35] Hartmann M, Schindler K, Gebbink T, Gritsch G, Kluge T. Pure
decomposition and multiset canonical correlation analysis. J EEG: Automatic EEG artifact removal for epilepsy monitoring.
Appl Math 2014. Neurophysiol Clin 2014;44:479—90.
[17] Cohen MX. Analyzing neural time series data: theory and [36] Hirsch L, Brenner R. Atlas of EEG in critical care. John Wiley
practice. MIT Press; 2014. p. 51—4. and Sons Ltd; 2010. p. 187—216.
[18] Daly I, Scherer R, Billinger M, Muller-Putz G. Force: [37] Hsu WY, Lin CH, Hsu HJ, Chen PH, Chen IR. Wavelet-based
fully online and automated artifact removal for brain- envelope features with automatic EOG artifact removal:
computer interfacing. IEEE Trans Neural Syst Rehabil Eng application to single-trial EEG data. Expert Syst Appl
2015;23:725—36. 2012;39:2743—9.
[19] De Clercq W, Vanrumste B, Papy JM, Van Paesschen W, Van [38] Hu J, Wang CS, Wu M, Du YX, He Y, She H. Removal of EOG and
Huffel S. Modeling common dynamics in multichannel signals EMG artifacts from EEG using combination of functional link
with applications to artifact and background removal in EEG neural network and adaptive neural fuzzy inference system.
recordings. IEEE Trans Biomed Eng 2005;52:2006—15. Neurocomputing 2015;151:278—87.
[20] De Clercq W, Vergult A, Vanrumste B, Van Paesschen W, Van [39] Islam MK, Rastegarnia A, Yang Z. A wavelet-based artifact
Huffel S. Canonical correlation analysis applied to remove reduction from scalp EEG for epileptic seizure detection. IEEE
muscle artifacts from the electroencephalogram. IEEE Trans J Biomed Health Inform 2016;20:1321—32.
Biomed Eng 2006;53:2583—7. [40] Jafarifarmand A, Badamchizadeh MA. Artifacts removal in
[21] De Vos M, Deburchgraeve W, Cherian P, Matic V, Swarte EEG signal using a new neural network enhanced adaptive
R, Govaert P, et al. Automated artifact removal as filter. Neurocomputing 2013;103:222—31.
304 M.K. Islam et al.

[41] James CJ, Gibson OJ. Temporally constrained ICA: an appli- filtering. In: Acoustics, Speech and Signal Processing (ICASSP),
cation to artifact rejection in electromagnetic brain signal 2012 IEEE International Conference. 2012. p. 661—4.
analysis. IEEE Trans Biomed Eng 2003;50:1108—16. [61] Nazarpour K, Wongsawat Y, Sanei S, Chambers JA, Oraintara
[42] Kierkels JJ, Riani J, Bergmans JW, Van Boxtel GJ. Using an S. Removal of the eye-blink artifacts from EEGs via STF-TS
eye tracker for accurate eye movement artifact correction. modeling and robust minimum variance beamforming. IEEE
IEEE Trans Biomed Eng 2007;54:1256—67. Trans Biomed Eng 2008;55:2221—31.
[43] Klados MA, Papadelis C, Braun C, Bamidis PD. Reg-ica: a hybrid [62] Ng SC, Raveendran P. Enhanced rhythm extraction using blind
methodology combining blind source separation and regres- source separation and wavelet transform. IEEE Trans Biomed
sion techniques for the rejection of ocular artifacts. Biomed Eng 2009;56:2024—34.
Signal Process Control 2011;6:291—300. [63] Nguyen HAT, Musson J, Li F, Wang W, Zhang G, Xu R, et al. EOG
[44] Lawhern V, Hairston WD, McDowell K, Westerfield M, Robbins artifact removal using a wavelet neural network. Neurocom-
K. Detection and classification of subject-generated artifacts puting 2012;97:374—89.
in EEG signals using autoregressive models. J Neurosci Meth- [64] Nicolaou N, Nasuto SJ. Automatic artefact removal from
ods 2012;208:181—9. event-related potentials via clustering. J VLSI Signal Process
[45] LeVan P, Urrestarazu E, Gotman J. A system for automatic Syst Signal Image Video Technol 2007;48:173—83.
artifact removal in ictal scalp EEG based on independent com- [65] Niedermeyer E, Da Silva FL. Electroencephalography: basic
ponent analysis and Bayesian classification. Clin Neurophysiol principles, clinical applications, and related fields. 5th edi-
2006;117:912—27. tion Lippincott Williams & Wilkins; 2004.
[46] Li Y, Ma Z, Lu W, Li Y. Automatic removal of the eye blink [66] Nolan H, Whelan R, Reilly R. Faster: fully automated sta-
artifact from EEG using an ICA-based template matching tistical thresholding for EEG artifact rejection. J Neurosci
approach. Physiol Meas 2006;27:425. Methods 2010;192:152—62.
[47] Lopez-Calderon J, Luck SJ. EEPLAB: an open-source toolbox [67] Noureddin B, Lawrence PD, Birch GE. Time-frequency anal-
for the analysis of event-related potentials. Front Hum Neu- ysis of eye blinks and saccades in EOG for EEG artifact
rosci 2014;8:1—14. removal. In: Neural Engineering, 2007. CNE’07. 3rd Interna-
[48] Ma J, Bayram S, Tao P, Svetnik V. High-throughput ocular arti- tional IEEE/EMBS Conference on neural Engineering. 2007. p.
fact reduction in multichannel electroencephalography (EEG) 564—7.
using component subspace projection. J Neurosci Methods [68] Noureddin B, Lawrence PD, Birch GE. Online removal of eye
2011;196:131—40. movement and blink EEG artifacts using a high-speed eye
[49] Ma J, Bayram S, Tao P, Svetnik V. Muscle artifacts in multi- tracker. IEEE Trans Biomed Eng 2012;59:2103—10.
channel EEG: characteristics and reduction. Clin Neurophysiol [69] Oostenveld R, Fries P, Maris E, Schoffelen JM. Fieldtrip: open
2012;123:1676—86. source software for advanced analysis of MEG, EEG, and
[50] Mahajan R, Morshed BI. Unsupervised eye blink artifact invasive electrophysiological data. Comput Intell Neurosci
denoising of EEG data with modified multiscale sample 2010;2011:1—9, http://dx.doi.org/10.1155/2011/156869
entropy, kurtosis, and wavelet-ICA. IEEE J Biomed Health [Article ID 156869].
Inform 2015;19:158—65. [70] O’Regan S. Artefact detection and removal algorithms for EEG
[51] Mallat S. A wavelet tour of signal processing: the sparse way. diagnostic systems. PhD thesis. University College Cork; 2013.
3rd edition Academic press; 2008. p. 535—90. [71] O’Regan S, Faul S, Marnane W. Automatic detection of EEG
[52] Mammone N, Morabito FC. Enhanced automatic wavelet artefacts arising from head movements using EEG and gyro-
independent component analysis for electroencephalographic scope signals. Med Eng Phys 2013;35:867—74.
artifact removal. Entropy 2014;16:6553—72. [72] Park HJ, Jeong DU, Park KS. Automated detection and elim-
[53] Maragos P, Kaiser J, Quatieri T. On amplitude and frequency ination of periodic ECG artifacts in EEG using the energy
demodulation using energy operators. IEEE Trans Signal Pro- interval histogram method. IEEE Trans Biomed Eng 2002;49:
cess 1993;41:1532—50. 1526—33.
[54] Mateo J, Torres AM, Garcia MA. Eye interference reduction [73] Peng H, Hu B, Shi Q, Ratcliffe M, Zhao Q, Qi Y, et al. Removal
in electroencephalogram recordings using a radial basic func- of ocular artifacts in EEG-an improved approach combining
tion. IET Signal Process 2013;7:565—76. DWT and ANC for portable applications. IEEE J Biomed Health
[55] Mihajlovic V, Li H, Grundlehner B, Penders J, Schouten Inform 2013;17:600—7.
A. Investigating the impact of force and movements on [74] Peyrodie L, Gallois P, Boudet S, Cao H, Barbaste P, Szurhaj
impedance magnitude and EEG. In: Engineering in Medicine W. Evaluation of the AFOP/DAFOP method for automatic fil-
and Biology Society (EMBC), 2013 35th Annual International tering of EEGs of patients with epilepsy. J Clin Neurophysiol
Conference of the IEEE. 2013. p. 1466—9. 2014;31:152—61.
[56] Mihajlovic V, Patki S, Grundlehner B. The impact of head [75] Potamianos A, Maragos P. A comparison of the energy opera-
movements on EEG and contact impedance: an adaptive tor and the Hilbert transform approach to signal and speech
filtering solution for motion artifact reduction. In: Engi- demodulation. Signal Process 1994;37:95—120.
neering in Medicine and Biology Society (EMBC), 2014 [76] Rankine L, Stevenson N, Mesbah M, Boashash B. A non-
36th Annual International Conference of the IEEE. 2014. stationary model of newborn EEG. IEEE Trans Biomed Eng
p. 5064—7. 2007;54:19—28.
[57] Mognon A, Jovicich J, Bruzzone L, Buiatti M. Adjust: an auto- [77] Rashed-Al-Mahfuz M, Islam MR, Hirose K, Molla MKI. Artifact
matic EEG artifact detector based on the joint use of spatial suppression and analysis of brain activities with electroen-
and temporal features. Psychophysiology 2011;48:229—40. cephalography signals. Neural Regen Res 2013;8:1500.
[58] Molla MKI, Islam MR, Tanaka T, Rutkowski TM. Artifact sup- [78] Richman JS, Moorman JR. Physiological time-series analysis
pression from EEG signals using data adaptive time domain using approximate entropy and sample entropy. Am J Physiol
filtering. Neurocomputing 2012;97:297—308. Heart Circ Physiol 2000;278:2039—49.
[59] Mukhopadhyay S, Ray G. A new interpretation of nonlinear [79] Roy V, Shukla S. Automatic removal of artifacts from EEG
energy operator and its efficacy in spike detection. IEEE Trans signal based on spatially constrained ICA using Daubechies
Biomed Eng 1998;45:180—7. wavelet. Int J Modern Educ Comput Sci (IJMECS) 2014;6:31.
[60] Navarro X, Poree F, Carrault G. ECG removal in preterm [80] Sameni R. The Open Source Electrophysiological Toolbox
EEG combining empirical mode decomposition and adaptive (OSET). http://www.oset.ir/ [Online].
Scalp EEG artifacts 305

[81] Sameni R, Gouy-Pailler C. An iterative subspace denoising [97] Ting K, Fung P, Chang C, Chan F. Automatic correction of arti-
algorithm for removing electroencephalogram ocular arti- fact from single-trial event-related potentials by blind source
facts. J Neurosci Methods 2014;225:97—105. separation using second order statistics only. Med Eng Phys
[82] Savelainen A. Movement artifact detection from electroen- 2006;28:780—94.
cephalogram utilizing accelerometer. Master’s thesis. Aalto [98] Turnip A. Automatic artifacts removal of EEG signals
University School of Science and Technology; 2011. using robust principal component analysis. In: Technol-
[83] Schetinin V, Schult J. The combined technique for detection ogy, Informatics, Management, Engineering, and Environment
of artifacts in clinical electroencephalograms of sleep- (TIME-E), 2014 2nd International Conference. 2014. p. 331—4.
ing newborns. IEEE Trans Inf Technol Biomed 2004;8: [99] Turnip A, Junaidi E. Removal artifacts from EEG signal using
28—35. independent component analysis and principal component
[84] Seneviratne U, Mohamed A, Cook M, D’Souza W. The analysis. In: Technology, Informatics, Management, Engi-
utility of ambulatory electroencephalography in routine clin- neering, and Environment (TIME-E), 2014 2nd International
ical practice: a critical review. Epilepsy Res 2013;105: Conference. 2014. p. 296—302.
1—12. [100] Vaughan TM, Heetderks W, Trejo L, Rymer W, Weinrich M,
[85] Shao SY, Shen KQ, Ong CJ, Wilder-Smith EP, Li XP. Auto- Moore M, et al. Brain-computer interface technology: a
matic EEG artifact removal: a weighted support vector review of the second international meeting. IEEE Trans Neural
machine approach with error correction. IEEE Trans Biomed Syst Rehabil 2003;11:94—109.
Eng 2009;56:336—44. [101] Wallstrom GL, Kass RE, Miller A, Cohn JF, Fox NA. Automatic
[86] Shoker L, Sanei S, Latif M. Removal of eye blinking artifacts correction of ocular artifacts in the EEG: a comparison of
from EEG incorporating a new constrained BSS algorithm. In: regression-based and component-based methods. Int J Psy-
Engineering in Medicine and Biology Society, 2004. IEMBS’04. chophysiol 2004;53:105—19.
26th Annual International Conference of the IEEE, volume 1. [102] Wang Z, Xu P, Liu T, Tian Y, Lei X, Yao D. Robust removal of
2004. p. 909—12. ocular artifacts by combining independent component anal-
[87] Shoker L, Sanei S, Chambers JA. Artifact removal from elec- ysis and system identification. Biomed Signal Process Control
troencephalograms using a hybrid BSS-SVM algorithm. IEEE 2014;10:250—9.
Signal Process Lett 2005;12:721—4. [103] Winkler I, Haufe S, Tangermann M. Automatic classification
[88] Skupch AM, Dollfuss P, Furbass F, Gritsch G, Hartmann MM, of artifactual ICA-components for artifact removal in EEG
Perko H, et al. Spatial correlation based artifact detection signals. Behav Brain Funct 2011;7:30.
for automatic seizure detection in EEG. In: Engineering in [104] Winkler I, Brandl S, Horn F, Waldburger E, Allefeld C,
Medicine and Biology Society (EMBC). 2013. p. 1972—5. Tangermann M. Robust artifactual independent compo-
[89] Stearns SD. Adaptive signal processing. Prentice Hall; 1985. nent classification for BCI practitioners. J Neural Eng
[90] Stevenson N, Rankine L, Mesbah M, Boashash B. Newborn EEG 2014;11:035013.
seizure simulation using time—frequency signal synthesis. In: [105] Yang Z, Liu W, Keshtkaran MR, Zhou Y, Xu J, Pikov V, et al.
Proc. APRS Workshop on Digital Image Computing. 2005. p. A EC-PC threshold estimation method for in vivo neural spike
145—51. detection. J Neural Eng 2012;9:046017.
[91] Sweeney K, Ward T, McLoone S. Artifact removal in physiolog- [106] Yong X, Ward RK, Birch GE. Artifact removal in EEG using
ical signals—practices and possibilities. IEEE Trans Inf Technol morphological component analysis. In: Acoustics, Speech and
Biomed 2012;16:488—500. Signal Processing, 2009. ICASSP 2009. IEEE International Con-
[92] Sweeney K, Ayaz H, Ward TE, Izzetoglu M, McLoone SF, ference. 2009. p. 345—8.
Onaral B. A methodology for validating artifact removal tech- [107] Yong X, Ward RK, Birch GE. Generalized morphological
niques for physiological signals. IEEE Trans Inf Technol Biomed component analysis for EEG source separation and artifact
2012;16:918—26. removal. In: Neural Engineering, 2009. NER’09. 4th Interna-
[93] Sweeney K. Motion artifact processing techniques for phys- tional IEEE/EMBS Conference. 2009. p. 343—6.
iological signals. PhD thesis. National University of Ireland [108] Zeng H, Song A, Yan R, Qin H. EOG artifact correction from
Maynooth; 2013. EEG recording using stationary subspace analysis and empiri-
[94] Sweeney K, McLoone SF, Ward TE. The use of ensemble empir- cal mode decomposition. Sensors 2013;13:14839—59.
ical mode decomposition with canonical correlation analysis [109] Zhao C, Qiu T. An automatic ocular artifacts removal method
as a novel artifact removal technique. IEEE Trans Biomed Eng based on wavelet-enhanced canonical correlation analysis.
2013;60:97—105. In: Engineering in Medicine and Biology Society, EMBC, 2011
[95] Takahashi T, Cho RY, Mizuno T, Kikuchi M, Murata T, Takahashi Annual International Conference of the IEEE. 2011. p. 4191—4.
K, et al. Antipsychotics reverse abnormal EEG complexity in [110] Zhao Q, Hu B, Shi Y, Li Y, Moore P, Sun M, et al. Automatic iden-
drug-naive schizophrenia: a multiscale entropy analysis. Neu- tification and removal of ocular artifacts in EEG — improved
roimage 2010;51:173—82. adaptive predictor filtering for portable applications. IEEE
[96] Teixeira AR, Tome AM, Lang EW, Gruber P, Da Silva AM. Trans Nanobioscience 2014;13:109—17.
Automatic removal of high-amplitude artefacts from single- [111] Zou Y, Nathan V, Jafari R. Automatic identification of artifact-
channel electroencephalograms. Comput Methods Programs related independent components for artifact removal in EEG
Biomed 2006;83:125—38. recordings. IEEE J Biomed Health Inform 2016;20:73—81.