biocybernetics and biomedical engineering 37 (2017) 401–411

Available online at www.sciencedirect.com

ScienceDirect

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

Original Research Article

Artifacts removal from EEG signal: FLM

optimization-based learning algorithm for neural
network-enhanced adaptive filtering

M.H. Quazi *, S.G. Kahalekar

Department of Instrumentation Engineering, Nanded, India

article info abstract

Article history: Electroencephalogram (EEG) denotes a neurophysiologic measurement, which observes the
Received 16 November 2016 electrical activity of the brain through making a record of the EEG signal from the electrodes
Received in revised form positioned on the scalp. The EEG signal gets mixed with other biological signals, called
11 March 2017 artifacts. Few artifacts include electromyogram (EMG), electrocardiogram (ECG) and elec-
Accepted 14 April 2017 trooculogram (EOG). Removal of artifacts from the EEG signal poses a great challenge in the
Available online 4 May 2017 medical field. Hence, the FLM (Firefly + Levenberg Marquardt) optimization-based learning
algorithm for neural network-enhanced adaptive filtering model is introduced to eliminate
Keywords: the artifacts from the EEG. Initially, the EEG signal was provided to the adaptive filter for
Electroencephalogram yielding the optimal weights using the renowned optimization algorithms, called firefly
Levenberg Marquardt algorithm algorithm and LM. These two algorithms are effectively hybridized and applied to the neural
Firefly algorithm network to find the optimal weights for adaptive filtering. Then, the designed filtering
Artifacts signal process renders an improved system for artifacts removal from the EEG signal. Finally, the
Signal to noise ratio performance of the proposed model and the existing models regarding SNR, computation
time, MSE and RMSE are analyzed. The results conclude that the proposed method achieves
a high SNR of 42.042 dB.
© 2017 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish
Academy of Sciences. Published by Elsevier B.V. All rights reserved.

are not only the pure form of the brain signal because some
1. Introduction
defected signals, called artifacts (such as, power line noise,
muscle contraction, heart activity and eye movement), also
Recently, a number of EEG-based applications, such as gets mixed with it [5,6]. Due to the existence of the artifact
wheelchair controllers and word speller programs [1–3] are signal, the examination of EEG signal from the EEG recordings
used by the researchers. Using the non-invasive measurement becomes more complex because it is perplexed with the
of EEG, the activities of the brain can be observed by placing the neurological patterns. To generate a correct analysis and
electrodes on the scalp of the human brain in multiple areas diagnosis [7], the unwanted signals must be eliminated from
[4]. The signal that results from recording the brain activities the recorded EEG signal.

* Corresponding author at: Department of Instrumentation Engineering, Nanded, India.

E-mail address: [email protected] (M.H. Quazi).
http://dx.doi.org/10.1016/j.bbe.2017.04.003
0208-5216/© 2017 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish Academy of Sciences. Published by Elsevier
B.V. All rights reserved.
402 biocybernetics and biomedical engineering 37 (2017) 401–411

Once the artifacts from the EEG signal were removed, the sources [24]. Additionally, the EEG signals are of very small
original brain signal can be extracted from the EEG recordings amplitudes, and because of that, they can be easily contami-
[4]. Though the removal of artifacts from the EEG signal is a nated by noise.
more complex process, it is utmost necessary for the progress The main difficulty of the EEG artifact removal problem is
of practical systems with low amplitude signal distortion [8]. the selection of the threshold level. The reason is that it should
Nowadays, more common methods are used to remove the not remove the original EEG signal coefficients and at the same
artifacts from the EEG signals. Some of them includes linear time, it should never keep the artifact signals as the original
filtering and regression model [9,10], ICA (independent ones.
component analysis) [11], WT (wavelet transform) [12–14] In the case of the time varying signals that arise from the
and (ANFIS) adaptive neuro-fuzzy inference system [7,15,16], human body, adaptive filtering is found to serve as an
adaptive filters and neural networks [3] and cascaded adaptive appropriate method for the EEG artifact removal. Although
filters [17]. ICA is one of the important analyses for sorting out the adaptive linear filters are the most widely used filters from
the EEG signals, which obtained from the electrodes that among the various available adaptive filters, their perfor-
positioned on the scalp, into a self-governing mechanism. mance is not satisfactory for dealing with the nonlinear
However, the ICA algorithm fails to yield better results during problems.
source separation. The artifacts signal was eliminated by the In [20], an adaptive FLN–RBFN-based filter was proposed to
PCA technique, which is used to handle the high dimensional, remove ocular, muscular and cardiac artifacts from the EEG
boisterous and concurrent data. However, the orthogonal signals. But, this adaptive neural network had not considered
rotation gets reduced by the PCA analysis. Another technique the weights optimally, when the learning process was
is the eye-blink artifacts removal by the adaptive filter performed. Hence, the finding of the optimal weights for
technique. Here, the adaptive filter subtracts the EEG source adaptive filtering is a heuristic search problem, and it should
signal from the estimated inference signal to remove the be performed optimally by handling all the constraints to
artifacts [18]. Also, the adaptive filters remove the artifacts in obtain a good artifacts removal performance.
real-time mode with less computation complexity. However,
no complete assurance can be given that the reference signal is
3. Adaptive noise cancelation for artifacts
a perfect signal. Investigation on the fault classification
removal
performance by the fuzzy logic techniques has also been
carried out [7].
The primary intention of this research is to design and Basically, adaptive noise cancelation is used to eliminate
develop a technique for removing the artifacts from the EEG the artifacts from the EEG signal. Fig. 1 shows the basic
signal. Here, we have planned to develop an FLM optimiza- block diagram representation for achieving adaptive noise
tion-based learning algorithm for the neural network-en- cancelation [20]. As shown in the figure, the adaptive
hanced adaptive filtering. This neural network-based noise cancelation process requires two inputs. The first
adaptive filtering performs the removal of artifacts from input is generated from the EEG signal source and it is
the EEG signals. At first, the EEG signal is given to the proposed represented as S(t). Then, the second input is collected from
adaptive filtering to obtain the optimal weights using the the source of the artifact signal and it is denoted as A(t). The
well-known optimization algorithm, called firefly algorithm noise source considered here represents the origin, where the
and LM (Levenberg Marquardt). These two algorithms were various artifacts such as, EOG, EMG and ECG generate. The
effectively hybridized and it is applied to the neural network noisy source signal of the artifacts signal gets passed through
to find the optimal weights for adaptive filtering. Then, the unidentified non-linear dynamics, resulting in the genera-
designed filtering was utilized for the removal of artifacts tion of the interference signal I(t). Then, the combination of
from the EEG signal. Finally, the comparison made against both the interference signal and the clean signal generates
ICA, wavelet ICA, Fast ICA, wavelet-based method that is the primary input signal and it is represented as follows:
given in [19] and the LM based optimization, regarding SNR,
percentage root mean square difference (PRD), and mean PðtÞ ¼ SðtÞ þ IðtÞ (1)
square error (MSE).
where S(t) is defined as the input source signal and I(t) indicates
the interference signal that is generated from the noise source.
2. Problem formulation Meanwhile, the signal generated from the noise source is applied
to an adaptive filtering process to get the filtered output. The
While recording the EEG signal through the biomedical filtered output is as close to the interference signal that is
equipment, a variety of large signal contaminations or noise generated from the result of nonlinear dynamics. The aim of
may affect the original signal data. The artifacts would this adaptive noise cancelation is to retrieve the clean EEG signal.
dramatically alter the signal that was recorded at all scalp In order to retrieve the clean EEG signal, the filtered output is
sites, especially at those sites that are closest to the source of subtracted from the primary input and it is shown below:
the noise. Hence, the artifact or noise cancelation serves as a
necessary stage of EEG processing [1,4,7,8,19–23]. So ðtÞ ¼ SðtÞ þ IðtÞFðtÞ (2)
One of the biggest challenges of using EEG is the very small
signal-to-noise ratio of the brain signals that we are trying to where F(t) denotes the filtered output and the output signal of
observe, due to the coupling of the wide variety of noise the entire adaptive noise cancelation is represented as So(t).
 biocybernetics and biomedical engineering 37 (2017) 401–411 403

Fig. 1 – Basic block diagram for adaptive noise cancelation.

is used to improve the optimization performance in this paper.

4. NARX system model Initially, the firefly algorithm considers as the learning
algorithm of the NARX neural network. The firefly algorithm
The NARX neural network model, which is defined as the depends on the blinking characteristics of the fireflies. The
combination of the multilayer feed forward network, recurrent possible solution vector in the firefly algorithm derives from
loop and time delay, is employed in this paper. This section the position of every firefly. Later, the position of every firefly is
shows the detailed description of the NARX neural network, changed based on the brightness. A firefly that does not find
which is used for modeling the nonlinear systems. Further- any brightness makes a random move. Here, the brightness
more, the NARX network is more suitable for modeling the value based on the performance that results from the
time series prediction analysis. Basically, the time series optimization problem. The random movement of the fireflies
response is based on the successively correlated signal, increases the computation complexity, in addition to taking
wherein the signal value depends on the past values and more time for converging at a solution. As a consequence, non-
the input signal. If the system input is found to be an feasible solutions and slower convergence result from the
assessable quantity, then that method can be considered as firefly algorithm.
NARX [25]. The NARX system model contains three vector To overcome these drawbacks of the firefly algorithm,
layers, namely, the input layer, the hidden layers and the a faster convergence nonlinear optimization algorithm, called
output layer. The input layer contains three kinds of collected the LM algorithm, is used along with the firefly algorithm. The
information vectors such as, exogenous input vector, delayed LM algorithm is used to find the location of the objects, which
regressed output vector and delayed exogenous input vector. can be defined as the difference between the investigational
After performing the neural network operation, the output values and the model equations. The LM algorithm is
vector is produced as L(n + 1). performed based on the combination of both the steepest
descent and the Gauss-Newton method. However, the LM
4.1. Internal architecture of the NARX neural network algorithm is used to perform a better-supervised learning with
high speed. So, the NARX neural network imparts a faster
Fig. 2 shows the internal architecture of the NARX neural training and more accuracy with the hybrid learning algo-
network. The mathematical form of the NARX neural work is rithm.
represented as follows: Fig. 3 shows the proposed hybrid learning algorithm
for the NARX neural network, in which three kinds of
Lðn þ 1Þ ¼ f ðLðnÞ; . . .; LðnDL Þ; VðnÞ; . . .; VðnDV ÞÞ (3) weights are combined using the learning algorithm. Here,
the weight input vector is generated from the weight of the
where L(n) is defined as the exogenous input vector. The exogenous input, regressed output and the functional
delayed values of the regressed output vector are represented results of both the exogenous as well as the regressed
as L(n  1), L(n  2), . . ., L(n  DL). Then, the series of delayed output. At first, the randomly generated input vector is
exogenous input vector is denoted as V(n), V(n  1),   , V given to the input of the both the LM and the FF
(n  DV). At the initial stage of the NARX neural network learning algorithm. After performing the optimization
operation, the weights assign to the exogenous input vector process, the weight vector is generated from both the
and the hidden units as well as between the regressed output Levenberg–Marquardt learning algorithm and the Firefly
vector and the hidden units. learning algorithm are WLM = Wl1 + Wl2 + Wl3 +    + Wlk and
Wff = Wf1 + Wf2 + Wf3 +    + Wfk respectively. Then, both of
4.2. LM-Firefly: proposed learning algorithm for the NARX these weight vectors are combined using the proposed
model hybrid learning algorithm, and a new weight vector
Whb = Wz1 + Wz2 + Wz3 +    + Wzk is generated.
A hybrid algorithm, developed by the combination of both the The steps that are involved in the hybrid learning algorithm
Firefly [26] and Levenberg–Marquardt (LM) algorithm [13], is discussed as follows:
404 biocybernetics and biomedical engineering 37 (2017) 401–411

Fig. 2 – Internal architecture of the NARX neural network.

Fig. 3 – Proposed hybrid learning algorithm for the NARX neural network.
 biocybernetics and biomedical engineering 37 (2017) 401–411 405

Step 1: Initially, assign the size of the population with where S(W) is defined as the value of the performance index, ET
dimension d. Additionally, make a random initialization of the denotes the target outputs and E refers to the simulated output.
values of both the attractiveness and the intensity factor. Step 10: Update the input weights based on the following
Moreover, initialize the fireflies randomly based on the weight equations of the LM algorithm.
vector. 1
The intensity of the Firefly is represented as follows: DW ¼ ½GT G þ hi GT E

where G is the Jacobian matrix, GT is the Jacobian transform

g0 matrix and h is the learning rate. The learning rate parameter
gd ¼
1 þ ld2 is updated using the decay function v. As the value of S(W)
where gd is defined as the intensity of the firefly at a distance d. increases, the learning rate gets multiplied by the decay func-
Then, the attractiveness of the firefly can be represented tion rate (v). Then, the value of S(W) is recomputed using the
as follows: weighted function W + DW as the trail weight. Similarly, the
learning rate is divided by the decay rate, whenever the func-
a0 tion S(W) increases. Then, the incremented values of the
ad ¼
1 þ ld2 weights is found by the formula,

Step 2: Evaluate the fitness function values. W ¼ W þ DW

Step 3: Calculate the brightness of every Firefly and
make a comparison between the fireflies based on the Step 11: Update the weight as per the following equation,
brightness value. Update the position of each firefly based when the trail-weighed function is less than the performance
on the brightness of the neighboring firefly. The weighted index.
function used for the firefly algorithm is represented
as follows: h ¼ hv

2 Here, the learning rate is updated with the decay value by

Wtþ1 ¼ Wt þ a0 eld ðWcb Wt Þ þ ge
ff
using the current learning rate. Then, return to Step 8 and
where the absorption coefficient of the light is denoted as l continue the process.
and e is a random number. Further, the current best solution is Step 12: Measure the learning rate as per the following
represented by Wcb. equation, when the performance index is less than the trail
Step 4: Once again, find the fitness value and the values of function.
attractiveness as well as the intensity. h
h¼
Step 5: Rank the fireflies to find the current best solution or v
the global solution.
Step 6: Continue this process, until the maximum number Step 13: Update the randomly generated weight vector
of iteration occurs. Once the firefly algorithm is applied, both using the LM formula, which is shown below:
the input vector and the updated weight are combined
as follows: LM
Wtþ1 ¼ Wt ½H þ m  T1  Q
Pff ¼ Vðw; Wnew Þ
where H is the Hessian matrix of the system and defined as the
new product of both the Jacobian matrix and the Jacobian trans-
where w is defined as the input vector and W is defined as
form matrix.
the weighted vector is generated from the firefly algorithm.

Step 7: Compute the error between the expected value and Step 14: Generate the gradient matrix from the Jacobian
the current value, as shown below. matrix and represent it as follows:
1X n
ff
Eff ¼ ðP Pgi Þ Q ¼ GT  E
n i¼1 i
where Q is the gradient matrix and is defined as the product of
ff
where Pgi denotes the original ground truth value and Pi both the Jacobian transform matrix and the error value.
defines the output of the firefly algorithm. Step 15: Represent the output of the LM algorithm based on
Step 8: Using the LM algorithm, initialize the weight vector the original input vector weight and newly updated weight,
based on the size of the hidden weights that are used in the which is expressed as follows:
NARX neural network, as shown below.
PLM
i ¼ Vðw; Wnew Þ
W ¼ Wa1 ; Wa2 ; . . .; Wak
Step 16: At the end of the LM algorithm, calculate the error
Step 9: Compute the sum of squared error based on the value using the updated weight and the ground truth value as
following equation. follows:

1X n

SðWÞ ¼ ET E ELM ¼ ðPLM Pgi Þ

n i¼1 i
406 biocybernetics and biomedical engineering 37 (2017) 401–411

where n is the total number of iterations, Pgi is the ground truth generating any interference signal such as EOG, ECG and EMG
value of the input vector and PLM
i is the vector that is generated signal.
from the output of the LM algorithm.
Step 17: Finally, generate the resultant weight vector
based on the results of both the firefly algorithm and the LM 5. Results and discussion
algorithm. These two algorithms have enabled optimization
in this paper. To perform the optimization process, the error In this paper, the real signal is obtained from the physionet
value of both the LM as well as the FF algorithm is database, and the performance analysis is carried out. For
considered for the analysis. When the error value of the demonstrating the proposed NARX model with the hybrid
LM algorithm is lesser than the error value of the FF learning method, a real signal is contaminated with various
algorithm (ELM < EFF), the weight vector is considered from artifact signal sources such as EOG, EMG, and ECG.
the result of the LM algorithm (Wtþ1 ¼ Wtþ1 LM
). Similarly, when
the error value of the FF algorithm is lesser than the error 5.1. Experimental setup
value of the LM algorithm (ELM > EFF), the weight vector is
selected from the result of the FF algorithm (Wtþ1 ¼ Wtþ1 FF
). As The proposed technique is implemented using MATLAB, and
a consequence, the current error value becomes lesser than the experimentation conduct with the real-time signals that
the previous error value (Et+1 < Et), and the value of the are available in physionet.
damping factor gets reduced. The value of damping factor
increases when the current error is greater than the 5.1.1. Dataset description
previous error (Et+1 > Et). Finally, the weight vector is The input signals are downloaded from the physionet by
generated through comparing the error value of the FLM accessing the URL (https://physionet.org/cgi-bin/atm/ATM?
learning algorithm. database=ptbdb&tool=plot_waveforms). Here, the input da-
tabase selected for the EEG is CHB-MIT scalp EEG database
4.3. Multi channel artifacts removal using LMF based (chbmit). From this EEG database, the initial five signals are
NARX model downloaded by giving the length as 1 min. For adding the
ECG artifact, the ECG signal is downloaded from apnea-ECG
Fig. 4 shows the block diagram representation of the NARX database (apnea-ecg) for the length of 1 min. Similarly, the
model that uses the MF-based learning algorithm to EMG signal is downloaded from the URL (https://physionet.
achieve multichannel artifacts removal. The electrical org/physiobank/database/emgdb/emg_healthy.txt). The EOG
activity of the brain can be measured using the electroen- signals are downloaded from the physionet by accessing
cephalogram method. In this method, 15 electrodes are the URL (https://physionet.org/cgi-bin/atm/ATM?database=
placed on the scalp of several areas of the brain. The signal, ptbdb&tool=plot_waveforms). Here, the input database
generated from the multiple electrodes, is defined as the selected for the EOG is UCD Sleep Apnea database (ucddb)
multi-channel signal. Then, the signal is generated from and the time length is 1 min. For adding the artifact signal
the multiple electrodes is passed through the NARX model such as, ECG, EOG and EMG with the EEG signal, the first
using the proposed LMF algorithm. The proposed NARX channel of the signal is considered as artifact signal. Then,
model generates the artifacts removed signal or clean EEG the artifact signal is added with all of the 15 channels of the
signal for the corresponding channel or electrodes, without five EEG signals taken for the experimentation. The

Fig. 4 – Block diagram representation to achieve multichannel artifacts removal using LMF-based NARX model.
 biocybernetics and biomedical engineering 37 (2017) 401–411 407

physiologic signals are sampled at 256 samples per second models like, ICA, WICA, FICA and cascaded adaptive filter that
with 16-bit resolution. takes values as 5.4679 dB, 5.8425 dB, 10.433 dB and 39.215 dB,
respectively. For the second signal, the SNR value of the ICA
5.2. Experimental results and the WICA model are 3.7654 dB and 4.0605 dB respectively,
and the SNR value of FICA is 8.2046 dB. The SNR value of
Various methods such as independent component analysis, cascaded adaptive filter is 34.299 dB. Here, the proposed hybrid
wavelet independent component analysis, fast independent optimization algorithm achieves the maximum SNR ratio of
component analysis, a neural network with LM, cascaded 38.359 dB. Once again, the SNR performance of the various
adaptive filter and the proposed NARX neural network are used models analyzes with the third signal source. Here, the SNR
for the extraction of the artifact signal from the input EEG signal. value attained by ICA and WICA are 5.4679 dB and 5.8425 dB.
The independent component analysis is used to reveal the The SNR value of FICA is 9.1664 dB, which is greater than the
hidden signal or the artifact signal from the input EEG signal. values of both the ICA and the WICA models and the SNR value
Also, the maximally independent source signals can be statisti- of cascaded adaptive filter is 33.712 dB.
cally estimated using the independent component analysis Fig. 6(b) shows the signal to noise ratio analysis of the EMG
approach. The estimation of an independent component artifacts signal that is presented in the EEG. By analyzing Fig. 6
analysis is performed one by one based on their negentropy, (b), the SNR values of the various artifacts removal models
which is used to perform a faster computation than the obtained. For the first signal, the SNR value of the ICA and the
independent component analysis. The wavelet decomposition WICA is 5.4348 dB and 5.5671 dB, respectively. The SNR value
algorithm is used to combine the wavelet transform with the obtained by the cascaded adaptive filter is 39.34 dB while the
independent component analysis, so as to overcome the SNR value of the proposed model is 42.042 dB. For the second
disadvantages that the two techniques impart separately. EEG signal applied to the EMG artifacts removal process, both
Fig. 5 shows the waveform representation of the ECG the NN_LM and the proposed model have achieved the same
artifact. The physiological artifacts such as ECG, EOC, and EMG SNR value as 38.397 dB, and the SNR value of the cascaded
arise from a variety of body activities. The artifacts signal of adaptive filter is 35.124 dB. Finally, we observe that the
electrocardiography (ECG) caused by the heart beats, which proposed model could effectively remove the EMG signal
generate the rhythmic activity in the EEG signal. Also, the artifact of EEG using the FLM learning algorithm.
responses of skin such as sweating may vary the electrodes' Fig. 6(c) shows the Signal to Noise Ratio analysis of the EEG
impedance and generate the ECG signal, in place of the signal with EOG artifacts. For the first signal, the SNR values
expected EEG signal. are 5.7113 dB, 5.7118 dB, 6.002 dB, and 39.293 dB for the
artifacts removal models, such as ICA, WICA, FICA, and
5.3. Performance analysis cascaded adaptive filter respectively. On the other hand, SNR
value obtained by the NN_LM and NN_FLM model is 42.042 dB.
5.3.1. Performance analysis based on signal to noise ratio From Fig. 6(c), we understand that the proposed method
In this section, the performance of the proposed model is reaches the higher SNR value, which achieves better perfor-
analyzed based on the signal to noise ratio, as shown in Fig. 6. mance of artifact removal.
Signal to noise ratio is one of the performance measurements,
which is used to detect the noise level in the original signal. 5.3.2. Performance analysis based on mean square error
Fig. 6(a) shows the SNR-based analysis that depends on the value
ECG artifacts signal of the EEG signal. In the first signal source, Fig. 7(a) shows the performance analysis of the proposed
the SNR value of the proposed model is 42.007 dB. This SNR method, involving the ECG artifact signal, using mean square
value is very high, when compared to the SNR of other existing error. The mean square error (MSE) of the WICA and NN_LM

Fig. 5 – ECG artifact cancelation process.

408 biocybernetics and biomedical engineering 37 (2017) 401–411

Fig. 6 – Signal to noise ratio analysis of EEG signal with various artifacts.

model is very high when compared to the proposed model. The above, we observed that the proposed model achieves
proposed model achieves as minimum mean square error minimum MSE value to improve the performance of the
value of 4882.6. Meanwhile, the maximum mean square artifacts removal process.
error that is obtained by the ICA model is 6915.9. Minimum
value of the mean square error results in a better performance 5.3.3. Performance analysis based on root mean square error
of the artifacts removal process. The proposed optimization- (RMSE) value
based learning algorithm also achieves the minimum mean Fig. 8 shows the RMSE analysis of the EEG signal with various
square error for removing the artifacts from the input EEG artifacts such as ECG, EMG, and EOG. Fig. 8(a) shows the RMSE
signal. analysis to assess the performance of the proposed model,
Fig. 7(b) shows the EMG artifacts signal removal process involving the ECG artifact signal. Usually, the minimum value
based on the mean square error performance. Initially, the of the RMSE generates a better performance. The five signal
MSE value of the various models such as ICA, WICA, FICA, sources are taken from the physionet database. The RMSE
NN_LM, cascaded adaptive filter and the proposed NN_FLM attained by the models like, FICA, NN_LM, and the cascaded
model are analyzed from the various signal source. When adaptive filter are 63.982, 0.26736 and 19.613 respectively
analyzing the first signal source, the proposed NN_FLM-based whereas, the value of RMSE obtained by the proposed model is
method achieves the minimum MSE value of 4310. Other 0.23922. From the figure, the least RMSE value is obtained by
existing models such as ICA, WICA, FICA, NN_LM, and the proposed model, when compared to the other existing
cascaded adaptive filter have obtained the maximum MSE models such as FICA and NN_LM and cascaded adaptive filter.
value as 5532.4, 5507.9, 5506.3, 4356.8, and 5156.1, respectively. Fig. 8(b) shows the performance analysis of the EMG signal
Fig. 7(c) shows the performance analysis based on the MSE artifact removal process based on the RMSE value. For the first
for the EOG signal removal from EEG. Here, the signals from a signal source, the proposed model obtains the RMSE value of
total of five patients considered. By analyzing the first patient's 0.22156, which is very less than the other existing models. The
brain signal, the MSE of models like ICA, WICA, and FICA are RMSE value of NN_LM model is 0.2232 and it nearly equals the
orderly obtained as 5508.2, 5502.4, and 5497.2. A lesser MSE RMSE of the proposed method. However, the proposed model
produces better performance. Furthermore, the NN_LM and obtains the least RMSE value. From the results, we understand
cascaded adaptive filter models have achieved the MSE value that the proposed model only renders a better artifacts
as 4667 and 5218.8. However, the proposed NN_FLM method removal performance, when compared to the other existing
has obtained the minimum MSE of 4369. From the results models.
biocybernetics and biomedical engineering 37 (2017) 401–411 409

Fig. 7 – Mean square error analysis.

Fig. 8 – RMSE analysis.

410 biocybernetics and biomedical engineering 37 (2017) 401–411

Table 1 – Comparative discussion of the proposed model implemented using MATLAB, and the experimentation has
with the existing models such as ICA, WICA, FICA, been conducted using the real-time signals that are available
NN_LM, and cascaded adaptive filter. in physionet. Here, 15 channel data are considered and the
MSE RMSE SNR artifacts like EOG, EMG, and ECG signals include for
the analysis. A comparison has been made subsequently
ICA [11] 5508.2 65.194 5.7113
WICA 5502.4 64.765 5.8425 using the independent component analysis, wavelet-indepen-
FICA 5492.6 63.982 10.433 dent component analysis, fast independent component
NN_LM 4356.8 0.2232 42.042 analysis and neural network-based LM algorithm, regarding
Cascaded adaptive filter [17] 5156.1 19.587 39.34 signal to noise ratio, root mean square error, mean square
NN_FLM (proposed) 4310 0.22156 42.042 error and computational time. From the results, we conclude
that the proposed FLM optimization-based learning algorithm
attains a better signal to noise ratio of 42.042 dB.
Fig. 8(c) shows the performance analysis based on RMSE for
the EOG artifact signal removal process. For the fifth signal references
source, the RMSE value of the various artifacts removal models
such as ICA, WICA, FICA, NN_LM, cascaded adaptive filter and
the proposed NN_FLM model are 85.709, 85.048, 85.038, 0.5889, [1] Jirayucharoensak S, Israsena P. Automatic removal of EEG
21.849, and 0.58697 respectively. From Fig. 8(c), the artifact artifacts using ICA and lifting wavelet transform.
removal models such as FICA, NN_LM and the proposed Proceedings of International Conference on Computer
NN_FLM model obtained the RMSE value of 82.501, 0.3737, and Science and Engineering; 2013. p. 136–9.
0.37008 respectively for the third signal source. Finally, we [2] Niedermeyer E, da Silva FL. Electroencephalography: basic
principles, clinical applications, and related fields.
conclude that the proposed model of NN_FLM has obtained
Lippincott Williams & Wilkins; 2005.
the lowest RMSE value that improves the performance of the
[3] Nguyen H-AT, Musson J, Li F, Wang W, Zhang G, Xu R, et al.
artifact removal process. EOG artifact removal using a wavelet neural network.
Neurocomputing 2012;97:374–89.
5.4. Comparative discussion [4] Turnip A, Junaidi E. Removal artifacts from EEG signal using
independent component analysis and principal component
Table 1 shows comparative discussion the of the proposed analysis. Proceedings of International Conference on
Technology, Informatics, Management, Engineering &
NN_FLM model with the existing models, such as ICA, WICA,
Environment; 2014. p. 19–21.
FICA, NN_LM, and cascaded adaptive filter based on MSE,
[5] Vazquez RR, Perez HV, Ranta R, Louis VD, Maquin D,
RMSE, and SNR. From the table, the value of MSE attained by Maillard L. Blind source separation, wavelet de-noising
the existing models, such as ICA, WICA, FICA, NN_LM, and and discriminant analysis for EEG artifacts and noise
cascaded adaptive filter are 5508.2, 5502.4, 5492.6, 4356.8, and cancelling. Biomed Signal Process Control
5156.1recpectively while, the proposed model achieves the 2012;7:389–400.
MSE value of 4310 which is smaller than the existing models. [6] Geetha G, Geethalakshmi SN. Artifact removal from EEG
using spatially constrained independent component
The value of RMSE obtained by the existing models such as
analysis and wavelet de-noising with Otsu's thresholding
ICA, WICA, FICA, NN_LM, and cascaded adaptive filter are technique. Procedia Eng 2012;30:1064–71.
65.194, 64.765, 63.982, 0.2232, and 19.587 respectively. The [7] FemilinSheniha S, SujaPriyadharsini S, Edward Rajan S.
proposed model achieves the RMSE value of 0.22156 which is Removal of artifact from EEG signal using differential
also smaller than the existing models. The minimum value of evolution algorithm. Proceedings of International
MSE and RMSE generates the better performance for removing Conference on Communication and Signal Processing; 2013.
p. 3–5.
the artifacts. The value of SNR obtained by the existing models
[8] Gorecka J, Walerjan P. Artifacts extraction from EEG data
such as ICA, WICA, FICA, NN_LM, and cascaded adaptive filter
using the infomax approach. Biocybern Biomed Eng 2011;31
are 5.7113 dB, 5.8425 dB, 10.433 dB, 42.042 dB, and 39.34 dB (4):59–74.
respectively whereas, the proposed model attains the higher [9] Correa G. Artifact removal from EEG signals using adaptive
SNR value of 42.042 dB. From the table, it is clear that the filters in cascade, vol. 90. IOP Publishing Ltd.; 2007. p. 1–10.
proposed model achieves the better performance while [10] Croft RJ, Barry RJ. Removal of ocular artifact from the EEG:
compared with the existing models. a review. Neurophysiol Clin/Clin Neurophysiol 2000;30
(1):5–19.
[11] Li Y. Automatic removal of the eye blink artifact from EEG
6. Conclusion using an ICA-based template matching approach. Physiol
Meas 2006;27:425–36.
[12] Senthilkumar P. Removal of ocular artifacts in the EEG
In this paper, a hybrid optimization algorithm for the neural through wavelet transform without using an EOG
network-enhanced adaptive filtering has been proposed to reference channel. Int J Open Probl Comput Sci 2008;1
remove the artifacts signal from the multi-channel EEG data. (3):188–200.
[13] Gaidhane VH, Singh V, Hote YV, Kumar M. New approaches
The proposed hybrid optimization algorithm has been used to
for image compression using neural network. Intell Learn
find the weights of the NARX neural network, and it has been Syst Appl 2011;3:220–9.
defined as the combination of both the LM as well as the FF [14] Krishnaveni V. Automatic identification and Removal of
algorithm, so as to overcome the individual shortcomings of ocular artifacts from EEG using wavelet transform. Meas Sci
the two algorithms. The proposed model of NN_FLM has been Rev 2006;6(4):45–57.
 biocybernetics and biomedical engineering 37 (2017) 401–411 411

[15] Suja Priyadharsini S, Edward Rajan S. An efficient soft- [21] Zou Y, Nathan V, Jafari R. Automatic identification of
computing technique for extraction of EEG signal from artifact-related independent components for artifact
tainted EEG signal. Appl Soft Comput 2012;12:1131–7. removal in EEG recordings. Biomed Health Inform 2014;20
[16] Liu T, Yao D. Removal of the ocular artifacts from EEG data (1):73–81.
using a cascaded spatio-temporal processing. Comput [22] Kim M-K, Kim S-P. Artifact removal from EEG signals using
Methods Prog Biomed 2006;83:95–103. the total variation method. Proceedings of International
[17] Correa AG, Laciar E, Patiño HD, Valentinuzzi ME. Artifact Conference on ASCC; 2015. p. 1–4.
removal from EEG signals using adaptive filters in cascade. [23] Turnip A. Automatic artifacts removal of EEG signals using
J Phys Conf Ser 2007;90(1):1–10. robust principal component analysis. Proceedings of
[18] Hsiao-Lung C, Yu-Tai T, Ling-Fu M, Tony W. The removal of International Conference on Technology, Informatics,
ocular artifacts from EEG signals using adaptive filters Management, Engineering & Environment; 2014. p. 19–21.
based on ocular source components. Ann Biomed Eng [24] Guruva Reddy A, Narava S. Artifact removal from EEG
2010;38(11). signals. Comput Appl 2013;77(13).
[19] Akhtar MT, Mitsuhasini W, James CJ. Employing spatially [25] Menezes Jr JMP, Barreto GA. Long-term time series
constrained ICA and wavelet de-noising, for automatic prediction with the NARX network: an empirical
removal of artifacts from multichannel EEG data. Signal evaluation. Neuro Comput 2008;71:3335–43.
Process 2012;92:401–16. [26] Subhani Shareef SK, Rasul Mohideen E, Ali L. Directed
[20] Jafarifarmand A, Badamchizadeh MA. Artifacts removal in Firefly algorithm for multimodal problems. Proceedings of
EEG signal using a new neural network enhanced adaptive International Conference on Computational Intelligence
filter. Neurocomputing 2013;103:222–31. and Computing Research; 2015.