Biomedical Signal Processing and Control 31 (2017) 199–210

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Real-time ocular artifacts removal of EEG data using a hybrid ICA-ANC

approach
Aysa Jafarifarmand a , Mohammad-Ali Badamchizadeh a,∗ , Sohrab Khanmohammadi a ,
Mohammad Ali Nazari b , Behzad Mozaffari Tazehkand a
a
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
b
Cognitive Neuroscience Laboratory, Department of Psychology, University of Tabriz, Tabriz, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Removal of ocular artifacts (OA) in real-time is an essential component in electroencephalography (EEG)
Received 16 March 2016 based brain computer interface (BCI) applications. However, many proposed artifact removal methods
Received in revised form 13 July 2016 are not applicable in real-time applications due to their time-consuming process. In this paper we propose
Accepted 12 August 2016
a hybrid approach based on a new combination of independent component analysis (ICA) and adaptive
noise cancellation (ANC). A particularly new feature of the proposed approach is the utilization of the
Keywords:
ICA decomposition to extract the artifact source signal to be used in ANC based on neural networks. The
Adaptive noise cancellation (ANC)
method performs using a few EEG signals without requiring any additional electrodes (e.g. electroocu-
Electroencephalography (EEG)
Independent component analysis (ICA)
lography). We show that the proposed approach is capable of effectively reducing the ocular artifacts in
Ocular artifacts (OAs) a negligible time delay well applicable in real-time BCI. In order to achieve reliable results, the proposed
Real-time artifact removal approach is evaluated using data recorded during cue-based BCI. The efficacy of the proposed approach
Real-time brain computer interface (BCI) in both offline and online performances is compared to several state of the art methods. The results
demonstrate that the proposed approach outperforms the compared methods in terms of removal of OA
and recovery of the underlying EEG.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction used: 1) Adaptive noise cancellation (ANC) is a relatively fast

approach [6,8]. EEG signals are corrected by subtracting the inter-
REAL-TIME brain computer interface (BCI) enables the ference contribution that is estimated by identifying a model
extremely disabled people to interact with the external world between artifact source and the corresponding interference sig-
using just the brain power, bypassing the role of the muscles. Elec- nal. A common problem in using such a method is its requirement
troencephalography (EEG) is the most widely used non-invasive to separately measuring the artifact source signals coincide with
brain-imaging technique for real-time BCI applications, mainly the EEG measurements. The other concern associated with this
due to its fine temporal resolution, ease of use, portability and approach is also that the directly measured artifact source signals
low cost. Unfortunately, EEG signals are highly contaminated by (especially the ones close to the head i.e. ocular activities) may
artifacts of both biological and technical origins [1]. Biological be contributed by EEG activities that may cause in loss of some
artifacts originating from eye blink/movements (ocular artifacts, cerebral information. 2) Independent component analysis (ICA) is
OA), heartbeat, and muscle activities interfere inevitably in EEG another powerful algorithm introduced to separate and remove a
measurements, among them OA have high amplitude signals that wide variety of EEG artifacts [9–11]. ICA decomposes the measured
obscure the information available in the EEG recordings. Therefore, EEG data into their underlying independent artifactual and cere-
removal of the OA is an essential component prior to the EEG bral components. The artifactual independent components (ICs) are
signals analysis. then rejected and the artifact-free EEG is reconstructed using the
A variety of methods have been proposed for EEG artifact residual ICs. Due to the large number of EEG channels required for
removal [2–11] among them two methods are the most widely precise separation of the components, ICA is not suited for portable
applications and also is a computationally demanding and there-
fore time consuming approach. The detection of the artifactual ICs
is also another time consuming step necessary in ICA based arti-
∗ Corresponding author. fact removal. The other concern about the ICA based algorithms is
E-mail address: [email protected] (M.-A. Badamchizadeh).

http://dx.doi.org/10.1016/j.bspc.2016.08.006
1746-8094/© 2016 Elsevier Ltd. All rights reserved.
200 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210

that the ICA models are applied for EEG analysis with underlying The particular feature of the proposed method is applying ICA
assumption that the cerebral and the artifactual sources are mixed decomposition just for extraction of the artifact source component
linearly, which might not always be true especially for electrodes to be utilized in ANC. Accordingly, the combination of the artifac-
located away from the signal origin. tual components is assumed to be linear just in the close distances.
In addition to the acceptable elimination level of the arti- This also allows us to benefit the advantage of NNs in adaptively
facts while preserving the cerebral information, a suitable artifact adjusting the parameters in order to well follow the changes. More-
removal method for a real-time BCI application should satisfy cer- over utilizing only a small number of EEG channels, in addition to
tain specifications containing convenience, results in speeding up the ICA decomposition toward
introducing no restriction for an adaptive real-time procedure. Con-
sequently, presenting an appropriate work-flow, the approach well
• Automatic performance, a practical BCI needs to perform fully
removes OA in real-time appropriate in BCI application.
automatic without any expert interaction,
• Utilizing suitable number and position of electrodes, attaching
high number of EEG electrodes or the electrodes around the arti- 2. Methods and materials
facts origins during long-term EEG recording in BCI applications
causes in inconvenience for the subject, and 2.1. EEG recording
• Real-time processing, artifact removal process for a real-time BCI
cannot introduce unacceptable time delays [12–14]. All EEG data used in this paper are recorded (from healthy
adult subjects) using a Mitsar® amplifier and WinEEG® software.
Explaining the procedure and the purpose of our study, all the sub-
Mateo et al. have applied ANC using radial basis function (RBF)
jects gave their informed consent to participate in the experiment.
neural networks to reduce OA from EEG signals in [15]. Simultane-
EEG signals are recorded using EEG cap by 19 electrodes including
ously measured ocular activities using electrooculography (EOG)
A1 and A2 located based on the international standardized 10–20
has been utilized as artifact source signal. Mahajan et al. have used
system. The signals are recorded using monopolar channels refer-
ICA to denoise eye blink artifact from EEG data in [16]. They have
enced to linked ear lobes, with the ground electrode placed on AFz.
used modified multi-scale sample entropy along with kurtosis to
Electrode impedance is also maintained below 10 k. The record-
identify the eye blink atrifactual independent components resulted
ings additionally contain the simultaneous measurement of vertical
by ICA decomposition. Multi-resolution wavelet analysis has then
EOG for testing purposes too.
been applied to denoise these components instead of completely
The data are collected from 10 participants (age between 20 and
zeroing the ICs in order to better retaining of the neural activities.
40 years). Toward ensuring the accuracy of the results, the record-
The algorithm is applicable for dense EEG systems and also is not
ings are done in two separate sessions for each subject. Recording
capable for real-time BCI applications.
procedure is designed based on the cue-based BCI paradigm applied
Peng et al. proposed an approach combining discrete wavelet
in BCI Competition 2008 [20]. The paradigm consisted of four
transformation (DWT) and ANC that removes only OA from EEG sig-
different motor imagery (MI) tasks, namely the imagination of
nals [6]. The method applies DWT to a single contributed EEG signal
movement of the left hand, right hand, both feet, and tongue.
in order to construct the OA source signal utilized in ANC removal
Recordings are done while the subjects are sitting in a comfort-
procedure. Using only one EEG channel without requiring the EOG
able armchair in front of a computer screen. At the beginning of
recordings, the method is suited for portable environments. How-
each session, a recording of approximately a minute of blinking is
ever, the performance accuracy is sensitive to the selection of
performed to be utilized for initial training of the algorithm for OA
wavelet basis and thresholding function that may cause in los-
removal. The procedure is then continued by classic cue-based BCI
ing the useful information or keeping the artifact interference. The
recording according to the arrow appears on the computer moni-
algorithm is also not suited to real-time BCI applications. Contin-
tor. All the data are recorded continuously in order to be applicable
uing the work, Zhao et al. have improved the process speed by
for the real-time process.
utilization of adaptive predictor filtering to recover true EEG by pre-
All EEG and EOG signals are recorded at sampling rate of 250 Hz
dicting EEG signal amplitudes in OA zones in [17]. The short-term
and filtered using a digital band-pass filter between 0.5 and 45 Hz
prediction constraint however restricts the real-time performance.
in real-time to especially suppress the line noise and also noise
The performance of the method is also heavily dependent on the
originating from the mains and preserve the data applicable for BCI
proper selection of the parameters.
application [12,14].
Breuer et al. have also proposed an algorithm based on con-
strained ICA (CICA) for real-time cardiac and ocular artifact
rejection in magneto-encephalography (MEG) in [18] and [19]. The 2.2. Positioning of ocular artifact source EEG electrodes
algorithm utilizes the knowledge derived from the simultaneously
measured artifact source signal (i.e. electrocardiogram, ECG or EOG) With respect to a practical BCI application, it is necessary to uti-
to optimize the cost function of infomax ICA. The algorithm is capa- lize the least possible number of EEG electrodes, which are also well
ble of real-time reduction of cardiac/ocular artifact, however, it positioned toward the subject’s convenience. Besides, the selected
requires large number of EEG channels for precise decomposition EEG channels should be well applicable to extract the pure arti-
and also additional ECG/EOG electrodes. fact source signal as possible. A tradeoff between the number of
In this paper, we propose a new fully automated hybrid the EEG electrodes and ICA decomposition quality should there-
approach, well suited for real-time applications applying very small fore be considered. For this purpose, the alternatives are selected
number of EEG electrodes, in order to remove OA from EEG signals. among the EEG electrodes with the most artifactual and the least
The method is based on combination of ICA and ANC techniques. cerebral activity contribution as possible. The more an EEG elec-
In the proposed method, we apply ICA decomposition to only a trode is close to the artifact origin, the more it is affected by the
few EEG signals close to the artifact origin. ANC based on a fully artifactual potential. This effect is also well accepted to be linear
automated neural network is then conducted using the resulted due to the close distance. Furthermore, EEG channels farther from
independent component (IC), the most relevant to the artifact, the motor area contain less neural contribution that may reduce
as the reference input bypassing the requirement to direct EOG the number of underlying independent components. Such choices
recording. consequently increase the ICA ability in a pure separation of the
 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210 201

low-pass Butterworth filter is preferred in order to have a smooth

filtering and well retaining of the desired frequency range.

2.3.2. Independent component analysis

ICA is a statistical technique to extract source signals from
their linear mixtures based on the assumption that the sources
are mutually independent. Assuming the independence of the arti-
fact and brain source signals that is well accepted due to their
independent origins, ICA has been introduced as a successful tool
in separating of variety of artifacts from EEG signals. Consider-
Fig. 1. Block diagram of the proposed ICA-ANC approach. ing X = (x1 , x2 , · · ·, xn )T as a set of n measured EEG signals, and
S = (s1 , s2 , · · ·, sn )T as underlying source signals, basic ICA model
can be written as the following linear relationship:
mentioned artifactual activity from the neural activities using the
few recorded EEG signals. Additionally, considering the fact that the X =A·S (1)
correlation between the resulted artifact source and the interfer-
where A is the unknown mixing matrix of dimension n × n. The
ence signal is the most important criterion for ANC to be executable,
difficulty in solving the problem is that neither the mixing matrix
EEG channels resulting in the OA source IC the most correlated with
nor the independent underlying components are known. ICA esti-
the directly measured EOG is preferred.
mates the independent components (ICs) by finding the right linear
In accordance with the fact that motor area, which controls
combinations of the mixture variables in order to maximizing the
the voluntary movements of skeletal muscles, is located immedi-
mutual statistically independence of the source signals such that
ately anterior to the central sulcus of the brain, the EEG electrodes
located distant from the central area are less contributed by the Ŝ = W · X (2)
motor imaginary cerebral tasks applied in BCI. Blinking and verti-
cal movements of the eyes (that are the case in the present paper) where Ŝ is the estimation of the ICs matrix S, and W is the estimated
also contribute the most in the far frontal sites (containing frontal unmixing matrix [22].
polar and frontal electrodes respectively) due to their location posi- EEG artifact removal using classic ICA is performed by elim-
tioned above the eyes [1,21]. We computed the correlation value inating the artrifactual ICs and remixing the remaining ones.
between the resulted OA source IC and EOG for several frontal EEG Considering the fact that the recorded EEG signals (particularly the
electrodes groups. The EEG channels group with the most corre- ones distant from the artifact origin) might not be the linear mix-
lation value was considered as the source channels for extraction ture of the mentioned artifact, in the present study we use just ICA
of the OA source signal. This was performed based on the multi- decomposition (utilizing the electrodes close to the artifact origin
ple analyses for different data segments of different subjects and as described in part B) to extract the artifact related IC that would
group of four channels Fp1, Fp2, F3, and F4 are reported as the suited be used as reference input for ANC.
channels group for extraction of the OA source IC. Hereinafter, the Several different ICA algorithms have been widely used for EEG
chosen EEG channels are referred to as OA source EEG data. signals source separation [23]. Here we use the Infomax algorithm
of Bell et al.[24]. The natural-gradient version of Infomax is utilized
[25],
2.3. Proposed ICA-ANC approach    
Wk =  I + 1 − 2y(Ŝk ) ŜkT Wk (4)
The block diagram of the proposed ICA-ANC approach is illus-
where I denotes the identity matrix,  is the learning rate, k corre-
trated in Fig. 1. The chosen OA source EEG signals are first filtered in
sponds to the current training time, and Ŝk is the intermediate ICs
order to eliminate the non-artifactual frequency range while pre-
matrix that is obtained by multiplying Wk by the matrix of mixtures.
serving the frequency content of the ocular artifact. Such filtering
The learning rate is reduced hierarchically during the training pro-
removes a large part of irrelevant sources in order to more accu-
cess in order to achieve an optimal solution point. y also represents
rate extraction of the OA source component through ICA. ICA is then
the sigmoid function that is the standard cost function applied in
applied to the filtered signals and the IC most related to the ocular
Infomax.
activity is considered for further analysis. Although, in the high-
In order to ensure the acceptable source component separation
density EEG recordings, it is possible to have more than one artifact
quality, it is necessary to apply ICA to adequate length epochs of EEG
related IC, due to the small number of EEG electrodes applied here,
data. In order to achieve a real-time analysis compatible approach,
the obtained results indicate that only one IC is the main source of
epochs of much less data samples are required. On the other hand, it
contamination. The derived IC and the intended EEG signal (the EEG
has been reported that the number of data samples more than a few
signal measured through the electrode from the intended channel
times the square of the number of the ICs is necessary for an optimal
position that is supposed to be cleaned in order to be applicable
ICA decomposition [18] that fortunately introduce no limitation,
in EEG applications such as BCI) are then entered respectively as
considering the small number of ICs to be estimated here. However,
reference and primary inputs to the ANC system. Identifying the
with accordance to the fact that the separation of the artifact source
dynamic function between the reference and primary inputs using
signal is meaningful only in the presence of the artifactual activity,
an adaptive filter (AF), the EEG signal is cleaned by subtraction of
it is important to take data segments with a length that ensures at
the resulted interference contribution. These steps are described in
least once occurrence of the artifact. The epoch length of EEG data is
detail in the following paragraphs.
therefore identified considering the blinking occurrence intervals.

2.3.1. Filtering 2.3.3. Automatic recognition of the artifact related component

The EOG artifacts generally have limited frequency content of Applying ICA decomposition to data segments containing at
below 15 Hz. Accordingly a filtering step is first conducted to the least one appearance of the ocular artifact, it is straight forward to
OA source EEG signals in order to remove the non-ocular activities recognize the OA related IC among the four obtained components.
as possible while keeping all the ocular activity intact. First order Considering the significantly high influence of the ocular activities
202 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210

Fig. 2. Structure of the FLN-RBFN based adaptive filter [31].

in the far frontal EEG electrodes rational to the neural activities,

the correlation between the signals of these electrodes and the OA
related IC is expected to be larger than that for other ICs [26,27].
Furthermore, unlike the motor imaginary tasks [14], the vertical
ocular activities interfere highly in the adjacent electrodes at both
left and right hemispheres. Therefore, similar to the metric in [28],
the mean of the absolute correlation (MAC) between each IC and the
EEG signals of the most affected channels Fp1 and Fp2 placed above
the eyes on opposite hemispheres is considered as the feature for
recognition of the OA related IC here.
 
MAC = 1⁄2 |(ŝj , x1 )| + |(ŝj , x2 )| (5)

where  (., .) represents the correlation between two signals,

ŝj (j = 1, 2, . . ., n) is the jth IC (jth row of matrix Ŝ), and x1 and x2
represent the EEG signals of channels Fp1 and Fp2. The IC with the
most resulted MAC value is taken as the OA source signal.

2.3.4. Adaptive noise cancellation

ANC is an algorithm to filter out additive interferences by identi-
fying a model between a measurable contamination origin and the
corresponding immeasurable interference, based on the assump-
tion that the interfered desired data is uncorrelated with the
interference while having an interference origin signal that is cor-
related with the interference and uncorrelated with the desired
data [8]. Considering the well accepted assumption of uncorrela-
tion between the cerebral EEG and artifact signals, ANC has been
introduced as a successful algorithm to remove artifacts from EEG
data, subject to accessing the artifact source signal. Directly mea-
suring of the artifact source signals, however, might be contributed
by the neural information that causes in loss of valuable EEG activ-
Fig. 3. The proposed modified training process for offline training of the FLN-RBFN
ities. Here we use the IC obtained from the previous level as the based adaptive filter.
reference input.
Assuming q (t) as the recorded sample of the intended EEG signal
at time t, and ŝind (t) as the sample of the OA related IC resulted by hybrid adaptive filter based on radial basis function neural net-
the ICA decomposition, which are fed to the canceller as primary works (RBFN) and functional-link neural network (FLN) [32] is
and reference inputs respectively, an adaptive filter is utilized to utilized owing to its particular features of automatic performance,
model the unknown dynamic function g between the artifact source and well capability of identification of nonlinear functions.
signal and interference contribution [29]. Corrected EEG c (t) is then The FLN-RBFN based filter is an expansion of RBFN that increases
achieved by subtracting the interference from the contaminated the dimension of the input space in the consequent part of the net-
intended EEG signal. work by applying FLN, using Chebyshev polynomials [33], in order
to enhance the ability of approximating nonlinear combinations of
c(t) = q(t) − g(ŝind (t)) (6)
the input variables.
Several types of adaptive filters have been proposed for arti- The structure of the adaptive filter for r inputs and Chebyshev
fact removal from EEG signals [30,31]. In this present study our polynomial of order m is shown in Fig. 2. It is seen that the output of
 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210 203

Fig. 4. (a) Schematic illustration of the proposed ICA-ANC based data processing procedure, (b) Real-time artifact removal algorithm.

the filter is the weighted average of the Chebyshev polynomials of Order of Chebyshev polynomial is considered equal to five in this
the inputs, instead of the weighted average of the inputs in classic step. Redundant neurons are then eliminated by considering the
RBFN. contribution of each existing neuron in reducing the total error
Here we made modifications in training process of the adaptive based on the obtained linear parameters.
filter in order to improve the performance toward the fully auto- e) Steps (a)–(d) are executed for all patterns of the data segment.
matic real-time application. Automatic training process applied for f) Obtaining the neurons based on all data samples, this step is
the FLN-RBFN based AF includes following steps (Block diagram is conducted in order to determine the adequate order of the
presented in Fig. 3): Chebyshev polynomial. For this purpose, Chebyshev polynomi-
als of order 3–7 are applied to compute the linear parameters
using LSE. The order resulted in the most mean square error
a) For each entering data sample, a new neuron is generated if (MSE) between the artifactual and cleaned EEG signals during
no neuron exists yet or none of the existing neurons satisfies the artifact occurrence is reported as the suitable value.
a predetermined minimum firing strength (MFS). g) Finally, the consequent linear parameters are calculated using
b) If no neuron is required to be generated for the sample, the mean least square error (LSE) based on the parameters achieved in
and width values of the existing neurons are allocated based on levels (a)–(f).
the new entering pattern in order to ensure the sufficient match
degree (i.e. MFS) for any pattern in the input space.
c) Steps (a) and (b) are repeated until the number of neurons is Optimizing the number of neurons based on the input patterns
equal to or less than three. provides the fast learning feasibility for the adaptive filter. Con-
d) If the number of generated neurons exceeds three, the conse- sequently, considering the low number of EEG signals utilized for
quent linear parameters are estimated based on the mean square ICA decomposition and also the fact that ANC based on the FLN-
cost function, using least square error (LSE) computation, and RBFN adaptive filter is a fast algorithm, the ICA-ANC based process
error reduction ratio (ERR) is calculated for all existing neurons. is fast enough to be applied in real-time clinical applications. The
204 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210

workflow of the proposed real-time process suited for online BCI For the adaptive FLN-RBFN based filter also the mentioned
applications is elucidated in the next section. length of data is initially used to determine the appropriate
parameters containing number of neurons and order of chebyshev
polynomials as explained in part C. With accordance to the nearly
2.4. Workflow of real-time artifact removal algorithm constant amplitude pattern of ocular activities, the derived num-
ber of neurons is an optimal alternative to be applied throughout
In order to have a real-time BCI applicable performance, it is the artifact removal. Thus, training process during the real-time
proposed to conduct the training and cleaning processes in parallel. data processing is then applied by the fixed number of RBF neu-
Fig. 4(a) presents the workflow of the real-time removal of an arti- rons. Additionally, again due to the overlap of the successive sliding
factual contribution from an EEG signal. The rather computationally windows the parameters calculated from the previous adaptive fil-
demanding processes of training ICA and the AF are performed as ter, AFi−1 , are applied as initial ones for estimating the current AFi .
separate task (Fig. 4(a), bottom line) in parallel to the cleaning pro- The consequent linear parameters are optimized using the recur-
cess (Fig. 4(a), up line). The cleaning process therefore consists of sive least square error (RLS) computation applying the parameters
a few simple computations (see Fig. 4(b)) adequate for real-time obtained from previous training as initial value. The considered
performance. overlap time for AF training however is less than that needed for
The removal procedure is performed using a sliding window of ICA due to overtraining issue.
T seconds with appropriate overlap. While calculating the actual
ICA unmixing matrix, namely Wi on the data segment i, and the 2.5. Performance evaluation
AF parameters, referred to as AFi , the latest computed values Wi−1
and AFi−1 are used for data cleaning. This way, the cleaning pro- 2.5.1. Compared approaches
cess bypasses the time needed for training. In the training process, We compare our proposed algorithm to our previous work and
after estimating Wi based on the filtered data, and then determin- three other state-of-the-art approaches.
ing the IC relevant to the artifactual activity (OA), the AFi is then
trained based on the derived IC segment and the contaminated 1) In our previous work [32], ANC based on the FLN-RBFN adap-
intended EEG signal of the same timeframe. In the cleaning pro- tive filter has been applied for online removal of the artifacts by
cess, (Fig. 4(b)), artifact related IC is extracted by multiplying only directly measuring the artifact source signals (e.g. EOG). This is
the predetermined row of called ANC in the rest of the paper.
Wi by the filtered artifact source EEG signals. The component is 2) The other approach is based on DWT and ANC [6]. Combined
then passed through the AFi , and the resulted signal is subtracted DWT-ANC approach applies DWT to a contaminated EEG to
from the contaminated EEG. Cleaning algorithm is executed for derive the artifact source signal to be used in ANC processing. The
each L new samples of recorded data, by considering the delay approach has utilized Daubechies 4 wavelet as mother wavelet
acceptable for intended BCI system. for removal of ocular artifacts. Seven layers of decomposition
In pursuance of the real-time proficiency, the goal is to achieve have been selected, and soft threshold has been applied to the
a high speed data cleaning without loss of the accuracy. In order to three lowest coefficients. We here utilize EEG channel Fp1 as the
perform the artifact removal using the nearest possible value to the most OA contaminated one, and also apply the same FLN-RBFN
actual time ICA unmixing matrix and AF parameters, it is important filter (instead of RLS algorithm used in the paper) for ANC in
to estimate the values as fast as possible. Therefore, both ICA and AF order to have a fair comparison. We call the method as WANC.
are initially trained and prepared based on an appropriate length 3) Ocular and cardiac artifact rejection for real-time analysis
of data, Tint , containing several times occurrence of the artifactual (OCARTA) is another algorithm that uses a modified version
activity at the first level. Using such individual based starting point, of Infomax ICA to remove OA and cardiac artifact in real-time
the ICA-ANC can then be trained in a few iterations in the next [18,19]. In the modified version, for OA removal, the sigmoid cost
coming steps. Consequently, the algorithm may require relatively function of Infomax is replaced by an estimated cumulative den-
longer time at first level for a rather high iteration training prior to sity function (cdf) of individual EOG signals that are measured
the real-time performance. simultaneously together with the EEG. This method is simply
Additionally in ICA training, by considering the overlap between referred to as ICA.
two consecutive data segments in a suited range, the underlying 4) The last approach is wavelet-ICA, in which EEG data are first
mixing process and accordingly the unmixing matrix is expected decomposed to underlying independent components by ICA.
to remain, to some degree, unchanged. Such appropriate over- The ICs distinguished as artifacts are then cleaned using dis-
lap range is determined by applying ICA decomposition to several crete wavelet transform instead of zeroing the ICs [16]. Modified
different data segments successively. Taking the suited overlap, multiscale sample entropy and kurtosis have been used to iden-
therefore, the latest calculated unmixing matrix Wi−1 from the pre- tify the OA related ICs, and fourth level of decomposition with
vious training serves as an optimal initial matrix for estimating Wi Biorthogonoal mother wavelet have been applied for denoising
at the ith segment that likewise reduces the training time in sub- using the soft threshold.
sequent trainings. Furthermore, the overlap and accordingly the
similar weight matrix brings the same dynamic for the underlying 2.5.2. Evaluation criteria
components. Therefore, the row that represents the ocular activity The success rate of an artifact removal algorithm depends on two
in Wi−1 must represent the ocular activity in Wi too. The identifi- main criteria that how well the artifacts were removed, and how
cation of the artifact related IC is, thus, performed just once after well the neural data were preserved. We utilize several metrics in
the first ICA decomposition. The same index, ind, is then consid- time and frequency domains to measure these criteria in order to
ered throughout the real-time process. Initial training is executed evaluate the OA removal results of the proposed ICA-ANC based
as mentioned in part C. After the first training, an adequately low algorithm, in both offline and online performance.
learning rate is taken for ICA in order to have a smooth learning In order to evaluate the performance in preserving the neu-
toward a mild update of weights. According to the high similarity ral information, originally artifact-free portions of EEG signals are
of the sequential unmixing matrices, high valued learning rate may utilized. These portions should remain intact after removal pro-
unnecessarily distance the weights from the optimal value, and so cess. The artifact-free portions are detected by an EEG expert. In
lengthen the training time. the time domain, we apply correlation coefficient, and exterior
 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210 205

standard deviation (STD) measure introduced in [34] between the Table 1

Results of Dynamical Stability Test of ICs.
selected EEG data portions before and after correction. Exterior STD
is defined as the mean squared difference between the STD of the Overlap Time Correlation
original and corrected signal over the uncontaminated portions. 11 s 0.99
We also calculate the root mean square error (RMSE) between the 10 s 0.98
signals. Correlation coefficient measures the similarity in shape of 9s 0.96
the waveforms, while exterior STD indicates the difference of the 8s 0.93
7s 0.91
powers of the signals [34]. RMSE measures both waveform and
6s 0.86
amplitude distortions. The closer the correlation coefficient value 5s 0.83
to 1, and the exterior STD and RMSE to 0 are, the better preservation 4s 0.78
of the neural information is achieved. Dynamic stability test of ocular related IC by computing the correlation between
In the frequency domain also, we utilize the mean absolute the ICA unmixing matrix weights. Values are averaged values resulted by testing
error (MAE) [17] that presents the frequency difference between over several subsequent data segments of 12 s length. At the overlap of 10 s the
the original and corrected signals in artifact-free portions. MAE correlation value is sufficiently large.

between the measured EEG signal q and corrected signal c is given

by
intervals between the successive blinks are measured on randomly

ω2 selected segments of random subjects that resulted in the mean of
MAE = |Pq (n) − Pc (n)|/(ω2 − ω1) 3.8 s and max of 8 s. Accordingly, data segment of (1.5 × 8 = ) 12 s is
n=ω1
considered in order to guarantee the presence of at least one blink-
ing occurrence (see Section 2). The same data segment of 12 s is also
Where Pq (n) and Pc (n) are power spectral densities (PSD) of orig- considered for training of the AF along the training process. Firstly,
inal artifact-free and corrected EEG signals, respectively, and ω1 an appropriate overlap percent between successive data segments
and ω2 are limits of the interested frequency band. is determined for training the ICA-ANC system along the real-time
In order to evaluate the performance in removal of ocular arti- performance. The performance of each approach is then reported in
facts, features of the entire processed signal segments and the offline application. Finally, the efficacy of the approaches in online
original artifact-free portions in the segments are compared as pro- performance is presented.
posed in [34]. Here we use the signal segments each of one specific
MI task in order to reduce the effect of nonstationarity of the signals.
3.1. Dynamical stability test of ICs
In time domain, difference between the STD of the entire segment
and mean STD over artifact-free portions is applied as introduced
In order to determine the proper overlap percent for the train-
in [34]. This metric is referred to as total STD difference.
ing process, we conducted the ICA decomposition to the sequential
In the frequency domain also, we introduce a metric to measure
data segments with overlaps of multiples of a second. More sim-
the PSD distortion based on what is applied in [35]. We compute
ilarity between the dynamic of the underlying sources in two
the mean of the difference between the PSD of the entire processed
successive data segments more correlated the related weights
signal segments and the mean PSD over original artifact-free por-
vector (with consideration to the amplitude ambiguity of ICA). Con-
tions in them over frequency bands. The metric will be referred
sequently, more accurate decomposition will be achieved using
to as mean PSD difference. With considering to the fact that the
fewer iterations in ICA training considering that the weights esti-
low frequencies (band , <4 Hz) are hardly observed in short-time
mated for the previous segment is the initial weights vector in
intervals of artifact free portions, the value is calculated for each of
current training (see Section 2). For this purpose, the correlation
the four  (4–8 Hz), ˛ (8–13 Hz), ˇ (13–30 Hz), (≥30 Hz) bands.
between the weights vector related to OA related ICs are calculated
We moreover examine the frequency correlation, fc , between the
for several overlap ranges. The experiment is executed over sev-
entire measured and corrected EEG signal segments to determine
eral data segments of various subjects and the averaged results are
the influence of the OA removal on the spectrum of the EEG signals.
illustrated in Table 1. The results indicate that the 12 s segments
Correlation in frequency domain between the EEG signals before
with overlap more than 10 s has sufficiently similar dynamic. Also,
and after artifact removal can be interpreted equivalent to the cor-
the minimum possible overlap range depends on the time needed
relation of the signals in time domain before and after filtering [7].
for ICA and ANC training with accordance to the fact that the re-
The frequency correlation between contaminated (q) and corrected
training is only possible after completion of the previous training.
(c) EEG signals in the frequency range of ω1, ω2 is as follows
However, thanks to the few needed iterations, small number of

 ω2 the components in ICA and fast learning capability of the AF, this

ω2
 
ω2
fc = 0.5 · (c ∗ q̃ + c̃q∗ )/ c ∗ q̃ · c̃q∗ (10) introduces no restriction. The computation time for estimating the
unmixing matrix of ICA and training the AF using 12 s of data is
ω1 ω1 ω1
found to be 40 ms and 20 ms on average (on an Intel Core i5-2430M,
where c̃ and q̃ are the Fourier coefficients of c and q, and c ∗ and q∗ are 2.4 GHz, 4-GB RAM) respectively during the real-time performance.
the complex conjugations of c̃ and q̃ respectively. If the frequency Based on these results, the 90% overlap (=10.8 s) is considered for
content of the signal is unaffected, fc gets 0, otherwise it takes a updating the ICA unmixing matrix in the present simulations. Fur-
value between 0 and 1. fc equal to 1 represents the complete loss thermore, in order to prevent the overtraining issues, 50% overlap
of the frequency content. is preferred for training of AF.

3. Results 3.2. Offline evaluation

The efficacy of the proposed ICA-ANC based algorithm in offline An example of utilizing the proposed ICA-ANC approach for OA
and online performance is examined and compared to the four removal is displayed in Fig. 5. The frequency domain correlation
mentioned approaches in OA removal using the (2 × 10 =) 20 BCI between the measured and corrected EEG signals indicates that
based EEG datasets recorded as elucidated in Section 2. In order the low frequencies, which are related to the ocular activities, are
to have a meaningful ICA decomposition for extraction of OA, the suppressed while the high frequency area is well retained. In other
206 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210

100

-100
0 2 4 6 8 10 12
time (s)
100
Original EEG
50 Corrected EEG by ICA-ANC
0
-50
6 6.5 7 7.5
time (s)
1

0.5

0
0 5 10 15 20 25 30 35 40 45
frequency (Hz)

Fig. 5. Top line: the OA contaminated EEG and the corrected EEG by ICA-ANC in time domain. Middle line: detailed view of an OA affected part. Bottom line: Frequency
domain correlation between contaminated and corrected EEG signal.

Fig. 6. Comparison of the three best removal algorithms in offline performance, and the detailed display of an OA-free portion.

words, the approach preserves the non-artifactual underlying EEG tion with the original EEG in frequencies above 20 Hz. It means that
signal while removing OA. the method outperforms in retaining the neural information dur-
An example of EEG data segment (12 s) corrected by WANC, ing the OA removal portions too. WICA and WANC also well retain
WICA, and the ICA-ANC artifact removal approach, as the three the non-OA frequencies, and seem to be in the next level. WANC
best methods, is also illustrated in Fig. 6. The figure presents a additionally presents more loss in higher frequencies (over 30 Hz).
comparison between an OA contaminated EEG signal and its cor- ICA and ANC have the most loss of useful frequencies respectively.
rection using the three approaches. All approaches remove the OA In order to quantitatively evaluate the performance of the
to a certain extent, however, in the detailed view over an OA-free approaches in retaining the original EEG, correlation coefficient,
partition, it can be seen that corrected EEG using WICA and the exterior STD difference, RMSE, and MAE were calculated over the
proposed approach better overlaps with the original EEG along the several randomly selected artifact-free portions of EEG data mea-
non-affected interval. In other words, the approaches preserve the sured from each subject over all EEG channels (Table 2). It can
neural information in uncontaminated partitions of EEG data better be seen that WICA, WANC and ICA-ANC are superior in preserv-
than the other compared approaches. ICA and ANC presented the ing the shape of the EEG signal with average correlation coefficient
most distortion that is why are not illustrated in figure (in order to value of more than 99%. The three approaches with the low average
have higher resolution view). Fig. 7 also presents the performance value of exterior STD difference introduce also the least distortion
of the five approaches in frequency domain. The signal corrected by in energy of original unaffected EEG. The lowest RMSE between the
the proposed ICA-ANC method has the highest frequency correla- original and corrected EEG signal by WICA and ICA-ANC addition-
 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210 207

Fig. 7. Frequency domain correlation between original and corrected EEG signals, comparison of five methods.

Table 2
Results of Frequency-Domain and Time-Domain in Offline Analysis Using Five Methods.

Metric Frequency Band Methods

ICA ANC WICA WANC ICA-ANC

Correlation Coefficient (×10−2 ) – 94.19 ± 11.18 97.42 ± 2.55 99.01 ± 0.94 99.39 ± 0.84 99.2 ± 0.98
Exterior STD Difference (×10−2 ) – 28.23 ± 76.74 21.03 ± 37.98 2.53 ± 6.36 2.94 ± 4.72 3.17 ± 4.66
RMSE (×10−1 ) – 2.48 ± 2.16 3.21 ± 2.62 0.98 ± 0.92 1.52 ± 1.46 1.05 ± 1.01
MAE 0.5–45 Hz 0.37 ± 0.89 0.62 ± 0.63 0.09 ± 0.23 0.25 ± 0.56 0.11 ± 0.35
Total STD Difference (×10−2 ) – 5.29 ± 42.14 15.47 ± 25.55 −6.63 ± 18.83 −32.46 ± 65.18 −2.86 ± 18.75
Mean PSD Difference  4.41 ± 4.22 5.93 ± 3.20 3.69 ± 3.29 2.55 ± 5.53 3.41 ± 3.01
(×10−2 ) ˛ 2.1 ± 56.53 2.98 ± 6.06 1.86 ± 5.74 1.48 ± 6.46 1.32 ± 5.91
ˇ 0.17 ± 1.15 0.13 ± 1.10 0.04 ± 0.91 −0.11 ± 0.93 −0.03 ± 1.08
0.17 ± 0.23 0.13 ± 0.27 0.14 ± 0.26 0.16 ± 0.26 0.12 ± 0.19

The values are given as mean ± standard deviation.

ally confirms that the proposed approach is successful in retaining underlying EEG power. Low deviation about the mean also indicates
the neural information. WANC is in the next place in this met- that the approach is well adaptive with various data of different
ric. Finally in frequency domain MAE indicates the best preserving EEG channels of different subjects. Considering the high positive
result for WICA and our ICA-ANC approach respectively. Overall ICA value for ANC, it is interpreted that the method removes too much
and ANC are not that successful in preserving the neural informa- data power, which might be due to the directly measured EOG con-
tion. WICA and WANC are superior as expected, with accordance to taminating to neural information. Using wavelet to extract the OA
the fact that the more pure artifact source signal is detected in them source in ANC, it is expected WANC to reduce the data energy loss.
thanks to the wavelet transform. However, efficacy of the methods High negative value, however, indicates that not enough power has
depends highly on correct selection of the threshold in wavelet. been removed from contaminated portions. The resulted value also
ICA-ANC is consequently appeared to have the most successful depends extremely on wavelet threshold, which meanwhile affects
result by its fully automatic performance. the neural data preservation. ICA introduces better performance
The efficacy of the five approaches in removal of the OA while but tending to energy loss that again can be obviated using the
preserving the underlying EEG was then evaluated by calculating wavelet transform to remove more pure artifact related part of ICs
total STD difference and mean PSD difference between the cor- in WICA.
rected EEG data segments and the artifact-free portions contained In frequency domain also, considering the mean PSD difference,
in the same segments of the original EEG over all the EEG data the proposed method is appeared to be superior in high frequency
segments related to different MI tasks (Table 2). Considering the bands (˛, ˇ, and ). In the low frequency band , WANC introduces
total STD difference metric, it can well be seen that the proposed better performance. The results over all show that the proposed
ICA-ANC method introduces the best OA removal while preserving ICA-ANC approach can more effectively eliminate the OA effect
208 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210

Table 3
Results of Online Analysis Using Five Methods.

Method Total STD (×10−2 )

MEAN STD MIN ABS MAX ABS

ICA 6.95 6.48 1.91 16.03

ANC 12.59 1.49 10.42 13.63
WICA −115.21 125.78 −9.07 −273.35
WANC −141.85 118.14 −36.04 −298.04
ICA-ANC −1.63 2.81 2.57 −3.29

Total STD difference is calculated for data segments of 6 s with different OA affect
severity. Results are illustrated in mean, standard deviation, value with minimum
and maximum abstract.
Fig. 8. Exterior STD difference in online performance.

absolute amount over all examined data segments are presented.

while retaining the neural information over various MI tasks related
Wavelet based approaches are appeared to be highly sensitive to
data of all participated subjects.
the changes. ANC presents well adaptation with the changes, how-
ever, results in power loss in cerebral information. ICA presents
3.3. Online evaluation better performance in higher amplitude OA contaminations. It can
be seen that the proposed ICA-ANC based algorithm well adapts to
The online performance of the ICA-ANC based algorithm is eval- the changes and preserves its well performance in OA removal.
uated here, based on the proposed real-time workflow (see Section A comparison of online OA removal by the proposed ICA-ANC,
2). The initial recording of the blinking state mentioned in section II WICA, and WANC approaches is shown in Fig. 9. Detailed views
is firstly applied for initial training of the system before the starting over an OA-free and an OA-affected portions confirm the obtained
of the actual real-time procedure. Such initial training will intro- results.
duce only a short delay just at the beginning of the experiment,
before starting of the real-time process. After initial training, in first
and second training after start of the real-time process ICA and AF 4. Discussion
required 2–3 iterations to converge. In the following, one iteration
is enough for ICA-ANC system to well adapt itself. With consider- We introduced a new approach suited to ocular artifacts
ation to the real-time BCI literature [14], the artifact removal in removal from EEG signals for real-time BCI applications using very
real-time procedure is applied for every 120 ms of recorded data. small number of EEG electrodes. The proposed method is based
The cleaning process for a data epoch of 120 ms lasts 0.8 ms in on two well-known algorithms ICA and ANC, where the reference
average using the proposed algorithm. input for ANC is extracted using ICA decomposition instead of direct
In order to compare the online artifact removal efficacy of the measurement through EOG recordings. In other words, the princi-
proposed algorithm with aforementioned four approaches, we per- ple of the proposed ICA-ANC approach is ANC without requiring the
formed the similar real-time procedure using the cleaning and additional artifact source electrodes and just using a few EEG chan-
training processes in parallel but with using each ANC, ICA, WANC, nels. Moreover, in contrast to the essential assumption made in
and WICA. For ICA we used the real-time workflow proposed in existing ICA based artifact removal approaches, in which the inde-
[18,19] in which three iterations is conducted in first training pendent components are linearly mixed in all EEG sensors, in our
continued by one iteration in the following trainings. WICA was proposed approach just the EEG channels close to the artifact source
conducted by 7 iterations of ICA based on the results achieved in (eyes) are supposed to be affected linearly. In order to estimate the
[18,19]. ANC was also performed by about 5 iterations in primary contribution in farther EEG electrodes instead we utilized the abil-
levels followed by single iteration trainings then [32]. For training ity of our previously proposed FLN-RBFN based adaptive filter in
of AF in WANC we took the same procedure as ANC. estimating nonlinear functions.
In order to detect the efficacy of the approaches in retaining Toward the real-time performance, modified training process is
the neural information in online performance, we computed exte- introduced for the FLN-RBFN based AF that gives it the quick adap-
rior STD difference over several artifact-free portions between EEG tation ability with the changes in signals’ properties. Thanks to the
signals before and after online artifact removal. The unaffected por- small number of components, ICA decomposition also well features
tions of EEG were selected randomly among the portions detected the quick following capability. On this basis real-time work-flow is
by an EEG expert of various subjects. The results are averaged over presented estimating a new ICA unmixing matrix and AF parame-
all EEG channels of all subjects. Fig. 8 presents the mean and stan- ters are conducted in parallel to data cleaning process. We firstly
dard deviation of the obtained values. Considering that the metric conduct an initial training based on a data segment containing sev-
would never have the negative value, standard deviations are illus- eral presences of the artifacts. Considering the same dynamic of
trated in just the positive side. It can be seen that WICA and ICA-ANC adequately overlapped data segments, the previously computed Wi
outperforms the other approaches by the lowest mean values. How- and AFi is then utilized as the initial values for subsequent training
ever, the rationally high standard deviation for WICA reports the of ICA and AF in following. Additionally, considering a small learn-
weaker adaptation ability of the approach. ing rate for ICA after initial training and determining an adequate
To evaluate the online performance in OA removal also we structure for the AF on initial training data segment accelerates
calculated the total STD difference in various OA affected data the learning process. Computing the correlation between the OA
segments of 6 s of all subjects that are so selected to wholly be extraction ICA weights for overlapped data segments of 12 s, it is
related to the one MI task each. In order to represent the capabil- shown in Table 1 that the training time less than 2 s well satisfies the
ity of the approaches to adapt with changes in data we selected dynamical stability of the underlying ICs. Accordingly, the proposed
the data segments with different maximum amplitude levels of parallel training introduces no restriction in real-time performance,
OA contribution. The results are reported in Table 3. Mean, stan- and removal process containing simple computation introduces a
dard deviation, and the value with the minimum and maximum negligible time delay.
 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210 209

Fig. 9. Comparison of the three algorithms in online performance, and the detailed display of an OA-free (bottom left) and an OA affected (bottom right) portion.

Executing the removal process every 120 ms (×250 Hz = 30 The proposed algorithm is tested using real EEG signals recorded
samples) data segment as an instance. the proposed algorithm from different subjects during the cue based BCI imaginary tasks.
introduces the time delay of 0.8 ms. Consequently, regarding the It is demonstrated that the new algorithm is capable of sufficiently
report of Lauer et al. [36] that human notices delays being more removal of OA and preserving neural information in real-time and
than 0.2 s, between the intent of movement and the performance outperforms several state of the art algorithms.
of the under control system, our proposed algorithm is well applica-
ble in real-time BCI applications. The algorithm is also illustrated to
be well adaptive with the signal properties change during real-time References
procedure, which is essential in EEG applications.
The proposed algorithm is tested in 20 EEG datasets recorded [1] L. Sörnmo, P. Laguna, Bioelectrical Signal Processing in Cardiac and
from 10 subjects during the cue based BCI imaginary tasks. The Neurological Applications, Academic Press, Burlington, 2005.
[2] S.Y. Shao, K.Q. Shen, C.J. Ong, E.P. Wilder-Smith, L. Xiao Ping, Automatic EEG
results are compared to four state of the art methods contain- artifact removal: a weighted support vector machine approach with error
ing: ANC using NNs, OCARTA, DWT-ANC, and wICA in both offline correction, IEEE Trans. Biomed. Eng. 56 (February (2)) (2009) 336–344.
and online performances. It is demonstrated that the new algo- [3] N. Mammone, F. La Foresta, F.C. Morabito, Automatic artifact rejection from
multichannel scalp EEG by wavelet ICA, IEEE Sens. J. 12 (3) (2012) 533–542.
rithm sufficiently removes OA while preserving the neural signal [4] H.A.T. Nguyen, J. Musson, F. Li, W. Wang, G. Zhang, R. Xu, C. Richey, T. Schnell,
of interest as measured by several metrics in both time and fre- F.D. McKenzie, J. Li, EOG artifact removal using a wavelet neural network,
quency domains. WICA demonstrates better efficacy in preserving Neurocomputing 97 (2012) 374–389.
[5] N. Mammone, F.C. Morabito, Enhanced automatic artifact detection based on
the underlying EEG using some metrics, however, it shows weak independent component analysis and Renyi’s entropy, Neural Netw. 21 (7)
performance in following the changes during online analysis. ANC (2008) 1029–1040.
also features well adaptation however does not introduce satisfying [6] H. Peng, B. Hu, Q. Shi, M. Ratcliffe, Q. Zhao, Y. Qi, G. Gao, Removal of ocular
artifacts in EEG − an improved approach combining DWT and ANC for
performance since the measured source EOG is affected by neural
portable applications, IEEE J. Biomed. Health Inform. 17 (3) (2013) 600–607.
activations. [7] V. Krishnaveni, S. Jayaraman, L. Anitha, K. Ramadoss, Removal of ocular
artifacts from EEG using adaptive thresholding of wavelet coefficients, J.
Neural Eng. 3 (4) (2006) 338–346.
5. Conclusion [8] P. He, G. Wilson, C. Russell, Removal of ocular artifacts from
electro-encephalogram by adaptive filtering, Med. Biol. Eng. Comput. 42 (3)
(2004) 407–412.
It is important to detect and remove various biological arti- [9] A. Delorme, T. Sejnowski, S. Makeig, Enhanced detection of artifacts in EEG
facts from EEG signals in many EEG based applications containing data using higher-order statistics and independent component analysis,
NeuroImage 34 (4) (2007) 1443–1449.
BCI. For real-time BCI applications, it is critical for artifact removal
[10] M. Bagheri Hamaneh, N. Chitravas, K. Kaiboriboon, S.D. Lhatoo, K.A. Loparo,
method to not introduce an unacceptable time delay and also to uti- Automated removal of EKG artifact from EEG data using independent
lize a few EEG electrodes toward the subject convenience in long component analysis and continuous wavelet transformation, IEEE Trans.
time utilizations. In this paper we have introduced a new approach Biomed. Eng. 61 (6) (Jun. 2014) 1634–1641.
[11] C. Burger, D.J. van den Heever, Removal of EOG artefacts by combining
suited for OA removal from EEG signals in real-time applications wavelet neural network and independent component analysis, Biomed. Signal
using very small number of electrodes. The proposed method is Process. Control 15 (2015) 67–79.
based on two well-known algorithms ICA and ANC, where the ref- [12] Z.S. Hippe, J.L. Kulikowski, EEG signal processing for BCI applications, human -
computer systems interaction: backgrounds and applications 2, in: Advances
erence input for ANC is extracted using ICA decomposition instead in Intelligent and Soft Computing, Springer, Rzeszów, Polonia, 2012, pp.
of direct measurement through EOG recordings. ICA decomposi- 571–591.
tion is performed using 4 EEG electrodes suitable positioned for [13] B. Luzheng, F. Xin-an, L. Yili, ‘EEG-based brain-controlled mobile robots: a
survey,’, IEEE Trans. Hum.—Mach. Syst. 43 (2) (2013) 161–176.
satisfying performance. Moreover, utilizing the parallel cleaning [14] J. Long, Y. Li, H. Wang, T. Yu, J. Pan, F. Li, A hybrid brain computer interface to
procedure the algorithm is well capable to remove the artifacts in control the direction and speed of a simulated or real wheelchair, IEEE Trans.
real-time. Neural Syst. Rehabil. Eng. 20 (September (5)) (2012) 720–729.
210 A. Jafarifarmand et al. / Biomedical Signal Processing and Control 31 (2017) 199–210

[15] J. Mateo, A.M. Torres, M.A. Garc, Eye interference reduction in [25] S.-i. Amari, A. Cichocki, H.H. Yang, A new learning algorithm for blind signal
electroencephalogram recordings using a radial basic function, IET Signal separation, Adv. Neural Inf. Process. Syst. (1996) 757–763.
Proc. 7 (7) (2013) 565–576. [26] L. Shoker, S. Sanei, J. Chambers, Artifact removal from electroencephalograms
[16] R. Mahajan, B.I. Morshed, Unsupervised eye blink artifact denoising of EEG using a hybrid BSS-SVM algorithm, IEEE Signal Process. Lett. 12 (10) (2005)
data with modified multiscale sample entropy, kurtosis, and wavelet-ICA, 721–724.
IEEE J. Biomed. Health Inform. 19 (1) (2015) 158–165. [27] S. Sanei, J.A. Chambers, EEG Signal Processing, Wiley, Chichester, UK, 2008.
[17] Q. Zhao, B. Hu, Y. Shi, Y. Li, P. Moore, M. Sun, H. Peng, Automatic identification [28] S. Shi-Yun, S. Kai-Quan, O. Chong Jin, E. Wilder-Smith, L. Xiao-Ping, Automatic
and removal of ocular artifacts in EEG-improved adaptive predictor filtering EEG artifact removal: a weighted support vector machine approach with error
for portable applications, IEEE Trans. Nanobioscience 13 (June (2)) (2014) correction, IEEE Trans. Biomed. Eng. 56 (2) (Feb. 2009) 336–344.
109–117. [29] M. Rupp, A family of adaptive filter algorithms with decorrelating properties,
[18] L. Breuer, J. Dammers, T.P.L. Roberts, N.J. Shah, A constrained ICA approach for IEEE Trans. Signal Process. 46 (3) (1998) 771–775.
real-time cardiac artifact rejection in magnetoencephalography, IEEE Trans. [30] S. Puthusserypady, T. Ratnarajah, H adaptive filters for eye blink artifact
Biomed. Eng. 61 (February (2)) (2014) 405–414. minimization from electroencephalogram, IEEE Signal Process. Lett. 12 (12)
[19] L. Breuer, J. Dammers, T.P.L. Roberts, N.J. Shah, Ocular and cardiac artifact (2005) 816–819.
rejection for real-time analysis in MEG, J. Neurosci. Methods 233 (2014) [31] B. Noureddin, P.D. Lawrence, G.E. Birch, Online removal of eye movement and
105–114. blink EEG artifacts using a high-speed eye tracker, IEEE Trans. Biomed. Eng. 59
[20] C. Brunner, R. Leeb, G.R. Muller-Putz, A. Schlogl, G. Pfurstscheller, BCI (8) (Aug. 2012) 2103–2110.
competition 2008 − Graz dataset IIA, 2008. [32] A. Jafarifarmand, M.A. Badamchizadeh, Artifacts removal in EEG signal using a
[21] R.M. Rangayyan, Biomedical Signal Analysis: A Case-Study Approach, Wiley, new neural network enhanced adaptive filter, Neurocomputing 103 (2013)
New York, NY, USA, 2002. 222–231.
[22] A. Hyvärinen, E. Oja, Independent component analysis: algorithms and [33] G. Panda, D.P. Das, Functional link artificial neural network for active control
applications, Neural Netw. 5 (2000) 411–430. of nonlinear noise process, pp. 163–166.
[23] L. Albera, A. Kachenoura, P. Comon, A. Karfoul, F. Wendling, L. Senhadji, I. [34] J.W. Kelly, D.P. Siewiorek, A. Smailagic, J.L. Collinger, D.J. Weber, W. Wang,
Merlet, ICA-Based EEG denoising: a comparative analysis of fifteen methods, ‘Fully automated reduction of ocular artifacts in high-dimensional neural
Bull. Pol. Acad. Sci. Tech. Sci. 60 (3) (2012) 407–418. data, IEEE Trans. Biomed. Eng. 58 (3) (2011) 598–606.
[24] A.J. Bell, T.J. Sejnowski, An information-maximization approach to blind [35] N.P. Castellanos, V.A. Makarov, Recovering EEG brain signals: artifact
separation and blind deconvolution, Neural Comput. 7 (November (6)) (1995) suppression with wavelet enhanced independent component analysis, J.
1129–1159. Neurosci. Methods 158 (2) (2006) 300–312.