Content uploaded by Rajagopal Gayathri
Author content
All content in this area was uploaded by Rajagopal Gayathri on Nov 24, 2020
Content may be subject to copyright.
RESEARCH ARTICLE
Detection and classification of electroencephalogram
signals for epilepsy disease using machine learning
methods
Rajagopalan Srinath
1
| Rajagopal Gayathri
2
1
Department of Electronics and
Communication Engineering, Vel Tech
Rangarajan Dr. Sagunthala R&D Institute
of Science and Technology, Avadi,
Chennai, India
2
Department of Electronics and
Communication Engineering, Sri
Venkateswara College of Engineering,
Pennalur, Sriperumbudur, India
Correspondence
Rajagopal Srinath, Department of
Electronics and Communication
Engineering, Vel Tech Rangarajan
Dr. Sagunthala R&D Institute of Science
and Technology, Avadi, Chennai 600062,
India.
Email: drsrinathrajagopalan@gmail.com
Abstract
The electroencephalogram (EEG) signal plays a key role in the diagnosis of
epilepsy. This study describes an automated classification of EEG signal for the
detection of Epilepsy disease using soft computing methods. The proposed
method is comprised of three modules: (a) transformation, (b) feature compu-
tation, and (c) feature classifications. In the first module, the nonsubsampled
contourlet transform is applied on the EEG signal which decomposes the
signal into approximate and directional subbands. The decomposition is done
using nonsubsampled pyramid filter bank and nonsubsampled directional
filter bank respectively. Secondly, the statistical features are extracted from the
decomposed directional subbands using wavelet packet decomposition
method. Finally, these features are classified by adaptive neuro-Fuzzy infer-
ence system classification method, which classifies the EEG signal into either
focal or nonfocal signal. The proposed method is tested on a set of EEG signals
for validation. The average classification rate of the proposed EEG signal clas-
sification system is 99.4%. The proposed EEG signal classification methodology
achieves a sensitivity of 99.7%, a specificity of 99.7%, and an accuracy of 99.4%.
The results confirmed that the proposed method has a potential in the classifi-
cation of EEG signals and thereby could further improve the diagnosis of
epilepsy.
KEYWORDS
ANFIS, EEG, epilepsy diagnosis, NSCT, soft computing
1|INTRODUCTION
Epilepsy is the disease that affects the human brain due
to the disorder in neurological central system. The abnor-
mal activity of brain cells are triggered by this epilepsy
disease. Electroencephalogram (EEG) is currently used
for acquiring the abnormal activities that occur in the
brain. The epilepsy disease can be detected by analyzing
the abnormal activities in the observed signals
1
. The epi-
lepsy can be classified into two types: primary epilepsy
and focal epilepsy. The patients who are affected by pri-
mary epilepsy are 20% and that of the focal epilepsy is
80% as per the report of World Health Organization 2010.
The signals from the brain area are classified into either
focal or nonfocal. The signal that comes from the abnor-
mal brain cells is called as focal signal, and the signal that
comes from the normal brain cells is called nonfocal sig-
nal. Hence, the epilepsy disease is detected by differenti-
ating and detecting the focal signal from nonfocal signal
2
.
Certain signal processing techniques have been
Received: 1 April 2020 Revised: 30 July 2020 Accepted: 3 August 2020
DOI: 10.1002/ima.22486
Int J Imaging Syst Technol. 2020;1–12. wileyonlinelibrary.com/journal/ima © 2020 Wiley Periodicals LLC 1
developed for detecting the epilepsy disease by differenti-
ating the focal from nonfocal signals
3
. The severity analy-
sis of the focal signal is also required for the patient
before surgery. The severity may either be “early”or
“advance”and the surgery is needed for those patients
with “advance”severity level.
The decomposition of EEG signal will help the radio-
logist to detect the focal behavior of the signal through
the extracted features from each of the decomposed
subband layers. Empirical mode decomposition (EMD),
linear prediction decomposition, and nonlinear mode
decomposition are the conventional methods for
decomposing the EEG signal in order to detect the focal
and nonfocal EEG signals. The stationary property of the
EEG signals cannot be detected and diagnosed using
these decomposition methods
4
. Hence, this study uses
dual-tree complex wavelet transform for decomposing
the EEG signals in order to analyze the stationary behav-
ior of the signals. Figure 1a shows the normal activity of
brain using EEG signal in nonfocal mode and Figure 1b
shows the abnormal activity of brain in focal mode.
This article is organized as follows: Section 2 states
the conventional classification methods of EEG signals.
Section 3 proposes an automated methodology for the
classification of EEG signal using machine learning algo-
rithm and Section 4 discusses the experimental results
performed on an open-access data set EEG signals for
epilepsy disease detection. Section 5 concludes this
article.
2|LITERATURE SURVEY
Many automated EEG signal classification systems using
different approaches have emerged in recent years.
Among such studies, Madhavan et al
5
presented a system
for differentiating the focal signals from nonfocal signals
using deep convolutional networks (DCN) based on time-
frequency domain. Synchro squeezing transform (SST)
was applied on EEG signals for decomposing the signals.
Then, the time-frequency features were extracted from
the decomposed SST and these features were classified
using DCN classification algorithm. The authors obtained
a classification accuracy of 99% on Bern-Barcelona EEG
data set.
Rahul Sharma et al
6
utilized third-order cumulant
function for the automatic detection of focal EEG signals.
The features were extracted from the EEG signals using
locality sensitive discriminant analysis (LSDA), method
and then these features were classified using support vec-
tor machine (SVM) classification technique. The authors
obtained a maximum classification accuracy of 99% on
Bern-Barcelona EEG data set. Siddharth et al
7
developed
a method for discriminating the focal signals from non-
focal EEG signals using sliding mode singular spectrum
analysis method. The reconstructed component features
were computed from EEG signals and those features
were classified using radial basis function neural net-
work. The authors tested their proposed method on the
EEG signals of the Barcelona EEG data set and obtained
an average accuracy of 99.11%, an average sensitivity of
98.52%, and an average specificity of 99.7%. Rahul
Sharma et al
8
presented a bispectrum method in order to
extract 25 magnitude features from EEG signals. These
features were obtained by using LSDA method and then
the computed features were classified using SVM classifi-
cation approach. The authors used 10-fold cross valida-
tion approach on the EEG signals of the Barcelona EEG
data set and obtained a classification accuracy of 96.2%.
Gupta et al
9
implemented the EMD method on the
EEG signals. Then, Sharma-Mittal entropy features were
computed from each of the decomposed subbands. These
nonlinear features were classified using least-squares
SVM (LSSVM) classification approach and the authors
had obtained a maximum classification accuracy of
83.18% on the EEG signals available in Barcelona EEG
data set. Yuan et al.
10
detected seizure points in the EEG
signals using deep learning algorithms. The inter-
correlation and intracorrelation features between each of
the variation points in the EEG signals were computed,
and those correlation features were classified using multi-
view deep learning method. Fivefold cross validation
method was incorporated for testing the effectiveness of
the proposed EEG signal classification. The proposed
FIGURE 1 Electroencephalogram (EEG) signals: A,
Nonfocal; B, Focal [Color figure can be viewed at
wileyonlinelibrary.com]
2SRINATH AND GAYATHRI
method stated in this work obtained a classification rate
of 94.37% on the EEG signals from an open-access
data set.
Gupta et al
11
discussed the Fourier-Bessel series
expansion method for the automatic classifications of
EEG signals. This method produced finite and unique set
of coefficients from the decomposition stages. Further-
more, 17 nonlinear features were computed and classified
using SVM classification method. This method was tested
by using 10-fold cross validation algorithm on the
Barcelona EEG data set. Bhattacharyya et al
12
proposed
an automated EEG signal classification system using
empirical wavelet transform method. The central ten-
dency measure features were computed from EEG signals
and they were classified using LSSVM classification
method. This method achieved a maximum classification
accuracy of 82.53%, sensitivity of 81.6%, and specificity of
83.46% on the EEG signals, which were available in the
Barcelona EEG data set. Kolekaret al
13
introduced the
bivariate EMD for the classification of focal and nonfocal
signals. The authors tested their proposed method using
the Bern-Barcelona EEG data set and obtained a classifi-
cation accuracy of 87.5%. Vipin Gupta et al
14
developed
a flexible analytic wavelet transform method for the
classification of EEG signals. The EEG signals were
decomposed at scale 15 and the Stein unbiased risk esti-
mate features were computed from the decomposed coef-
ficients. The LSSVM classifier was used to classify the
extracted features and the authors had obtained an aver-
age classification accuracy of 94.41%. Truong et al
15
designed a deep learning algorithm based on con-
volutional neural networks for the classifications of EEG
signals. The authors had used Max-pooling algorithm
with 5 * 5 spatial filters in each of the convolutional
layers for optimizing the classifications.
Pushpendra singh et al
16
investigated the Fourier
series-based decomposition methodology for detecting
and classifying the EEG signals. The authors had con-
structed a linear random model for this purpose. It was
applied on 700 EEG signals out of which, 692 signals
were classified correctly. The authors had obtained a
detection rate of 98.8%. Sharma et al
17
demonstrated a
tunable-Q wavelet transform (TQWT) method in order to
obtain the decomposed signal coefficients. The nonlinear
features were computed from each of the decomposed
subband coefficients and later those features were classi-
fied using SVM classification approach. The authors had
obtained an average classification accuracy of 94.06% on
the signals of the Barcelona EEG data set. Bhattacharyya
et al
18
illustrated TQWT for decomposing the EEG sig-
nals. The authors had extracted multivariate fuzzy
entropy features from each of the decomposed subband
matrices and those features were classified using LSSVM
classification approach. This method achieved an average
classification accuracy of 84.67% on the open-access Bar-
celona EEG data set. Sharma et al
19
advocated an auto-
mated EEG focal signal classification using orthogonal
wavelet filter banks. The signals that were passed
through these wavelet filter banks were decomposed at
various levels of scaling factors. The obtained decomposi-
tion coefficients were further classified using LSSVM
classification approach and thereby the authors had
obtained a classification accuracy of 94.25%, a sensitivity
of 91.95% and a specificity of 96.56%. Dalal et al
20
had
explored the flexible analytic wavelet transform (FAWT)
for obtaining the decomposition coefficients of EEG sig-
nals. This nonstationary transform produced fractal
dimension features at each of the scaling levels. This
method was tested by using the Kruskal-Wallis statistical
test, and the authors had obtained an average classifica-
tion accuracy of 89.1% on the EEG signals of the Barce-
lona EEG data set.
Deivasigamani et al
21
analyzed a soft computing-
based adaptive neuro-Fuzzy inference system (ANFIS)
classification method for differentiating the focal signal
and nonfocal signals using their feature values. It was
applied on 700 EEG signals out of which 694 signals were
classified correctly. The authors had obtained a detection
rate of 99.1%. Abhinaya et al
22
had dealt with a method-
ology which extracted entropy-based features from the
EEG signals. Those extracted features were optimized
using sequential forward feature selection method. Then,
SVM in linear regression mode was applied on the opti-
mized features for differentiating the focal and nonfocal
EEG signals. The authors had obtained a classification
rate of 92.8% by applying it on the open-access EEG data
set. Sharma et al
23
had decomposed the EEG signals with
respect to their X- and Y-channel using EMD. This model
had determined the maxima and minima for each of the
variations in the signal, and their residue was computed
using the average value of the maxima and minima
points. The entropy and variance for each of the
decomposed components were computed and the signals
were classified using SVM in least square mode.
Sharma et al
24
had assessed the discrete wavelet
transform (DWT) and the features were based on
entropy. Those computed features were ranked using
Bhattacharyya space algorithm and the ranked features
were further classified using various machine learning
classification algorithms such as K-nearest neighbor and
probabilistic neural networks. Sharma et al
25
had
highlighted the EMD of EEG signals in order to obtain
the intrinsic mode functions. The entropy features from
each of the IMF were computed and they were further
classified using LSSVM classification approach. The
authors had obtained an average classification accuracy
SRINATH AND GAYATHRI 3
of 87% of the EEG signals of the Barcelona EEG data set.
Gupta et al
26
had evaluated the TQWT for obtaining the
decomposition filter coefficients from EEG signals. The
mixture entropy features were computed from each of
the decomposed subband coefficients and then these mix-
ture features were classified using SVM classification
approach which was based on radial basis functions. This
method had obtained a maximum classification accuracy
of 90.01%.
3|MATERIAL AND METHODS
In this study, shearlet transform-based random forest
(RF) classification approach is used for the classification
of focal and nonfocal EEG signals. Initially, the features
are extracted from the coefficients of the shearlet trans-
form. Then those features are optimized using genetic
algorithm and finally the optimized features classified
using RF classification approach. Figure 2 shows the pro-
posed EEG signal classification using RF classification
approach.
3.1 |Materials
This study uses the Bern-Barcelona EEG data set
27
for
the classification of EEG signals. This is an open-access
data set and was created in the year 2012, which was per-
mitted all the researchers to use the acquired EEG sig-
nals. The EEG signals in this data set are obtained from
five patients over a period of 80 hours at the Neurology
Department of the Bern University. This data set contains
7500 pairs of EEG signals and these signals are catego-
rized into focal (3750) and nonfocal EEG signal (3750).
All these EEG signals are quantized at the frequency
range of 512 Hz for a time duration of 20 seconds. Each
EEG signal in this data set is represented by its X- and
Y-channel, respectively. From this dataset, 750 focal and
750 nonfocal EEG signals are obtained for evaluating the
performance of the proposed method. The signals used in
this study are further spilt into training and testing data
set. The training data set contains 150 focal and 150 non-
focal EEG signals. The testing data set contains 350 focal
and 350 nonfocal EEG signals. The EEG signals in both
of the training and testing data sets are independent of
each other.
3.2 |Methods
The nonsubsampled contourlet transform (NSCT) is
applied on the EEG signal which decomposes the source
signal into approximate and directional subbands. The
features are extracted from those subbands and then they
are finally classified by ANFIS classification method.
Figure 2 shows the proposed methodology for the classifi-
cations of EEG signal.
3.3 |Nonsubsampled contourlet
transform
In this study, three-level decomposition is carried out on
EEG signals using NSCT. The NSCT consist of non-
subsampled pyramid filter bank (NSPF) and non-
subsampled directional filter bank (NSDF). First, the
EEG signals are passed through NSPF, which decom-
poses the signal into low-pass and high-pass signals. The
high-pass signals are passed through NSDF, which pro-
duces directional subbands. At decomposition stage
2, the low-pass subband signals are passed through NSPF,
which produces low- and high-frequency subbands. This
high-frequency subband signals are again passed through
NSDF for obtaining the directional subbands. This pro-
cess is repeated for the third-stage decomposition also.
Figure 3 shows the decomposition of EEG signals using
three-level NSCT.
In this study, MATLAB toolbox NSCT is used for
decomposing the EEG signals into approximation and
directional subbands.
3.4 |Feature extractions
The focal and nonfocal EEG signals can be differentiated
with the help of features that are extracted from the decom-
position subbands. Characteristic feature vector (CFR),
probability density function (PDF), mutual information
(MI) features, energy and pattern spectrum entropy (PSE)
features are extracted for the classification of EEG signals.
FIGURE 2 Proposed methodology for electroencephalogram
(EEG) signal classifications
4SRINATH AND GAYATHRI
3.5 |Characteristic feature vector
In this study, the CFR features were computed by
decomposing the EEG signals with the help of wavelet
packet decomposition (WPD) method. It is the extension of
wavelet decomposition (WD), in which the signals can be
passed through low-pass filter (LPF) g(n) and high-pass filter
(HPF) h(n) with decimation factor 2. The LPF produces
approximation subbands and HPF produces detailed sub-
bands, at level 1 of the decomposition. At level 2 of the
FIGURE 3 Decomposition
of electroencephalogram (EEG)
signals using three-level
nonsubsampled contourlet
transform (NSCT)
FIGURE 4 Third level
decomposition of
electroencephalogram (EEG)
signal: A, wavelet
decomposition of signal; B,
wavelet packet decomposition of
signal
SRINATH AND GAYATHRI 5
decomposition, the approximation subbands are again pas-
sed through LPF and HPF, which produces approximation
and detailed subbands. This process is repeated until level
3 coefficients are obtained. The main function of DWT is to
convert the signals from time domain representation into
frequency domain representation. The main limitation of
FIGURE 5 Level 3 decomposition of electroencephalogram (EEG) signal: A, wavelet packet decomposition (WPD) of X-channel;
(B) WPD of Y-channel [Color figure can be viewed at wileyonlinelibrary.com]
6SRINATH AND GAYATHRI
this filter is that its time resolution is reduced and frequency
resolution is increased at each of the consecutive stages of
decomposition. The produced frequency resolution bands
arenotofthesamesize.Thiswillcreatethefrequencyinsta-
bility on decomposed subbands as depicted in Figure 4a.
In order to overcome the limitation of WD, WPD is
used for decomposing the EEG signals, in which the sig-
nals passed through LPD and HPD produce approxima-
tion and detailed subbands, respectively. At the level 2 of
decomposition, the approximation and detailed subbands
are passed through LPF and HPF in order to produce
equal frequency subbands. This process is repeated until
level 3 coefficients are obtained as shown in Figure 4b.
Figure 5a shows the WPD of X-channel and Figure 5b
shows the WPD of Y-channel at the level 3 of the decom-
position stages.
From the level 3 subband decomposition of WPD, the
energy factors are computed for each of the subbands S
3j
using the following equation:
Ej=ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
7
j=0
S3jjðÞ
2
:
v
u
u
tð1Þ
Then, the Eigen vectors are computed using energy
factors of each of the individual subbands at level 3 with
the help of the following equation:
E=ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
7
j=0
Ej2
:
v
u
u
tð2Þ
The CFR is computed using Eigen vectors and indi-
vidual energy factors for each of the subbands at the level
3 of decomposition using the following equation:
CFR = Ej
E
j=0,…,7
:ð3Þ
Table 1 show the extracted CFR features that are used
to differentiate the focal and non-focal EEG signals.
3.6 |Probability density function
In this study, PDF features are computed for X- and Y-
channel of the EEG signals with respect to normal distri-
bution considering zero mean and unique SD as input
parameters.
The mean of EEG signal is computed as
μ=1
NX
N
i=1
xiðÞ,ð4Þ
whereas, Nis the total number of samples in EEG signal.
The SD of EEG signal is computed using the follow-
ing equation:
σ2=1
N−1X
N−1
i=0
xi2−
1
NX
N−1
i=0
xi
!
2
"#
:ð5Þ
The PDF of the EEG signal is computed as
PDF xðÞ=1
ffiffiffiffiffiffiffiffiffiffi
2πσ2
pe
−x−μðÞ
2
2σ2
:ð6Þ
Figure 6a,b shows the extracted PDF features for both
X- and Y-channels of the focal and nonfocal EEG signal,
respectively.
3.7 |PSE features
The shape-size complexity of the signal can be defined
with the help of PSE features. It can be computed using
the following equation:
PSE = −X
N
i=0
xiðÞlog xiðÞ:ð7Þ
Tables 2 and 3 show the PSE features of the focal and
nonfocal EEG signals, respectively.
TABLE 1 Characteristic feature vector (CFR) computation for focal and nonfocal electroencephalogram (EEG) signals
Features Focal EEG signal Nonfocal EEG signal
Energy factor for X-channel {7.56, 0.85, 0.45, 0.23} {9.49, 0.94, 0.50, 0.25}
Energy factor for Y-channel {5.91, 0.78, 0.42, 0.21} {6.15, 1.69, 0.88, 0.44}
Eigen vector for X-channel 95.43 105.84
Eigen vector for Y-channel 85.68 95.78
CFR for X-channel {79.29, 8.95, 4.76, 2.42} {89.74, 8.96, 4.73, 2.39}
CFR for Y-channel {69.06, 9.16, 4.94, 2.50} {64.24, 17.65, 9.21, 4.66}
SRINATH AND GAYATHRI 7
3.8 |MI features
The mutual information features are computed between
X- and Y-channel of the EEG signals. The MI between X-
and Y-channel of the EEG signals is computed using the
following equation:
MI 1 = Entropy XðÞ−Entropy YðÞ−Joint Entropy X,YðÞ:
ð8Þ
The MI between Y- and X-channel of the EEG signal
is computed using the following equation:
FIGURE 6 Probability density function (PDF) features of X- and Y-channel: A, focal signal; B, nonfocal signal [Color figure can be
viewed at wileyonlinelibrary.com]
8SRINATH AND GAYATHRI
MI 2 = Joint Entropy XðÞ−Entropy YðÞ−Joint Entropy X,YðÞ:
ð9Þ
Table 4 shows the MI features of focal and nonfocal
EEG signals.
3.9 |Energy features
The energy level of the quantitative sampling points in
the EEG signal represent the energy level of the signal
for the given time period. The focal and nonfocal EEG
signals can have different energy levels with respect to its
sampling point variations.
The energy feature of X-channel is computed as
E1= ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
N−1
i=0
x2
i:
v
u
u
tð10Þ
The energy feature of Y-channel is computed as
E2= ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
N−1
i=0
y2
i:
v
u
u
tð11Þ
Table 5 shows the energy features for focal and non-
focal EEG signals.
3.10 |Classifications
Initially, the ANFIS classifier is trained by the extracted
features of the focal and nonfocal signals. The trained pat-
terns are generated in the ANFIS training mode. Then, the
extracted features from the test EEG signals are classified
by ANFIS classification mode with respect to the trained
patterns. The classification mode of the ANFIS classifier
produces single output response, which may be either low
or high. The low value indicates that the test EEG signal
belongs to the focal category, and high value indicates that
the test EEG signal belongs to the nonfocal category.
4|RESULTS AND DISCUSSIONS
The proposed EEG classification methodology was simu-
lated using MATLAB R2014b simulation software. The
TABLE 2 Energy and pattern spectrum entropy (PSE) features
for focal electroencephalogram (EEG) signals
EEG signal
sequences
Focal EEG signal
PSE for
X-channel
PSE for
Y-channel
1 8.09 * 10
4
0.37 * 10
4
2 9.01 * 10
4
0.24 * 10
4
3 8.78 * 10
4
0.35 * 10
4
4 8.89 * 10
4
0.31 * 10
4
5 7.09 * 10
4
0.29 * 10
4
TABLE 3 Energy and pattern spectrum entropy (PSE) features
for nonfocal electroencephalogram (EEG) signals
EEG signal
sequences
Non- focal EEG signal
PSE for
X-channel
PSE for
Y-channel
1 13.13 * 10
4
5.6 * 10
4
2 12.10 * 10
4
6.1 * 10
4
3 13.06 * 10
4
6.3 * 10
4
4 12.98 * 10
4
5.9 * 10
4
5 12.56 * 10
4
5.8 * 10
4
TABLE 4 Mutual information (MI) features for focal and
nonfocal electroencephalogram (EEG) signals
Focal EEG signal Nonfocal EEG signal
MI1 MI2 MI1 MI2
3.2733 3.0444 3.3632 2.2348
2.6997 2.4782 2.7789 1.6541
2.3476 2.1320 2.4215 1.2962
2.0672 1.8554 2.1544 1.0257
1.8416 1.6275 1.9371 0.8137
1.6506 1.4447 1.7516 0.6435
1.4949 1.2920 1.5863 0.5064
1.3645 1.1641 1.4469 0.3947
1.2566 1.0562 1.3256 0.3078
1.1666 0.9657 1.2222 0.2389
1.0904 0.8887 1.1335 0.1863
1.0254 0.8242 1.0558 0.1488
0.9679 0.7646 0.9872 0.1211
0.9161 0.7138 0.9240 0.1037
0.8699 0.6677 0.8690 0.0908
0.8228 0.6249 0.8195 0.0843
0.7762 0.5861 0.7742 0.0835
0.7326 0.5495 0.7303 0.0855
0.6907 0.5176 0.6896 0.0907
0.6486 0.4898 0.6527 0.0955
0.6114 0.4623 0.6198 0.1011
SRINATH AND GAYATHRI 9
experiments were conducted using Intel Dual Core 2 pro-
cessor with 4 GB internal RAM. The proposed system
detected 349 focal EEG signals correctly over 350 focal
signals. Hence, the classification rate with respect to focal
signals is 99.7%. The proposed system also detected
347 nonfocal EEG signals correctly over 350 nonfocal sig-
nals. Hence the classification rate with respect to non-
focal signals is about 99.1%. The average classification
rate of the proposed EEG signal classification system
is 99.4%.
The following performance evaluation metrics were
also used to determine the performance of the proposed
EEG signal classification system:
Sensitivity SeðÞ=TP
TP + FNðÞ
,ð12Þ
Specificity SpðÞ=TN
TN + FPðÞ
,ð13Þ
Accuracy AccðÞ=TP + TN
TP + FP + TN + FNðÞ
,ð14Þ
Precision PrðÞ=TP
TP + FPðÞ
,ð15Þ
Positive Predictive Rate PPRðÞ=TP
TP + FPðÞ
,ð16Þ
Negative Predictive Rate NPRðÞ=TN
TN + FNðÞ
,ð17Þ
whereas TP is true positive, which is 349; TN is true nega-
tive, which is 347; FP is false positive, which is 1; and FN
is false negative, which is 3.
Table 6 shows the analysis of proposed focal and non-
focal EEG signal classifications. The proposed EEG signal
classification methodology achieved a sensitivity of
99.7%, a specificity of 99.7% an accuracy of 99.4% with
PPV as 98.6% and NPV as 97.9%.
Table 7 shows the comparisons of proposed EEG clas-
sification methodology with other state-of-the art
TABLE 5 Energy features for focal and nonfocal electroencephalogram (EEG) signals
EEG signal
sequences
Focal EEG signal Focal EEG signal
Energy features for
X-channel
Energy features for
Y-channel
Energy features for
X-channel
Energy features for
Y-channel
1 9.5 * 10
3
6.4 * 10
3
7.6 * 10
3
5.9 * 10
3
2 9.2 * 10
3
6.2 * 10
3
7.1 * 10
3
5.1 * 10
3
3 9.9 * 10
3
6.1 * 10
3
7.9 * 10
3
5.8 * 10
3
4 9.7 * 10
3
6.9 * 10
3
7.8 * 10
3
5.3 * 10
3
5 9.1 * 10
3
6.7 * 10
3
7.3 * 10
3
5.2 * 10
3
TABLE 6 Analysis of proposed focal and nonfocal
electroencephalogram (EEG) signal classifications
Performance metrics Simulation results (%)
Sensitivity 99.7
Specificity 99.7
Accuracy 99.4
Precision 99.3
Positive predictive rate 98.6
Negative predictive rate 97.9
Classification rate 99.4
TABLE 7 Comparisons of
proposed electroencephalogram (EEG)
classification methodology with other
state-of-the arts methods
Methodology
Total number of
EEG signals
Number of EEG signals
correctly classified
Classification
rate (%)
Proposed work
(this study)
700 696 99.4
Pushpendrasingh
et al
16
700 692 98.8
Deivasigamani
et al
21
700 694 99.1
Abhinaya et al
22
700 650 92.8
10 SRINATH AND GAYATHRI
methods. The proposed work stated in this study accu-
rately classified 696 signals over 700 signals and achieved
a classification rate of 99.4%. Pushpendrasingh et al
16
classified 692 signals accurately over 700 signals and
achieved a classification rate of 98.8% using Fourier
series-based decomposition methodology. Deivasigamani
et al
21
classified 694 signals accurately over 700 signals
and achieved a classification rate of 99.1% using ANFIS
classification method. Abhinaya et al
22
classified 650 sig-
nals accurately over 700 signals and achieved a classifica-
tion rate of 92.8% using SVM classification method. From
Table 7, it is very clear that the proposed methodology
had achieved a higher classification rate when compared
to other state-of-the art methods.
5|CONCLUSIONS
This article presented an EEG data classification method-
ology, based on WPD, makes the decision about the brain
state of the patient. This work uses NSCT for decomposing
the EEG signals and then features are extracted from the
decomposed subbands. Then, ANFIS classification algo-
rithm is used to classify the source EEG signal into focal
and nonfocal. The proposed methodology for EEG signal
classification system achieves higher accuracy (99.4%),
sensitivity (99.7%), and specificity (99.7%) with 99.4% of
classification rate. Therefore, the conclusion is that the
proposed method can be used to classify EEG signals in an
effective manner. The simulation results of the proposed
methodology are compared with conventional methodolo-
gies. There are future improvements that can be done by
implementing it using deep learning algorithms. Though
they will definitely result in a more complex analysis yet
will give us more accurate classification results for EEG
severity diagnosis system in future. Also the detection
capabilities should be identified by analyzing long-term
continuous EEG recordings.
CONFLICT OF INTEREST
The authors declare no potential conflict of interest.
ORCID
Rajagopalan Srinath https://orcid.org/0000-0002-3723-
9678
REFERENCES
1. Goshvarpour A, Ebrahimnezhad H, Goshvarpour A. Classifica-
tion of epileptic EEG signals using time-delay neural networks
and probabilistic neural networks. Int J Inf Eng Electron Busi-
ness. 2013;5(1):59-67.
2. Andrzejak RG, Schindler K, Rummel C. Nonrandomness,
nonlinear dependence, and non-stationarity of
electroencephalographic recordings from epilepsy patients.
Phys Rev. 2012;E86:046206.
3. Tiwari A, Pachori RB, Kanhangad V, Panigrahi B. Automated
diagnosis of epilepsy using key-point based local binary pattern
of EEG signals. IEEE J Biomed Health Inform. 2016;99:888-896.
4. Xiang J, Li C, Li H, et al. The detection of epileptic seizure sig-
nals based on fuzzy entropy. J Neurosci Methods. 2015;243:18-25.
5. Madhavan S, Tripathy RK, Pachori RB. Time-frequency
domain deep convolutional neural network for the classifica-
tion of focal and non-focal EEG signals. IEEE Sensors J. 2020;
20(6):3078-3086.
6. Sharma R, Sircar P, Pachori RB. A new technique for classifica-
tion of focal and non-focal EEG signals using higher order
spectra. J Mech Med Biol. 2019;19(01):1940010.
7. Siddharth T, Tripathy RK, Pachori RB. Discrimination of focal
and non-focal seizures from EEG signals using sliding mode
singular Spectrum analysis. IEEE Sensors J. 2019;19(24):12286-
12296.
8. Sharma R, Pachori RB, Gautam S. Empirical mode decomposi-
tion based classification of focal and non-focal EEG signals.
Proc Int Conf Med Biometrics. 2015;1:135-140.
9. Gupta V, Pachori RBA. New method for classification of focal
and non-focal EEG signals. Machine intelligence and signal
analysis. Advances in Intelligent Systems and Computing. Vol
748. Singapore: Springer; 2019.
10. Yuan Y, Xun G, Jia K, Zhang A. A multi-view deep learning
framework for EEG seizure detection. IEEE J Biomed Health
Inform. 2019;23(1):83-94.
11. Gupta S, Krishna KH, Pachori RB, Tanveer M. Fourier-Bessel
series expansion based technique for automated classification
of focal and non-focal EEG signals. International Joint Confer-
ence on Neural Networks (IJCNN), Rio de Janeiro. Rio de
Janeiro, Brazil: IEEE; 2018:1-6.
12. Bhattacharyya A, Sharma M, Pachori RB, Sircar P,
Acharya UR. A novel approach for automated detection of
focal EEG signals using empirical wavelet transform. Neural
Comput Appl. 2018;29:47-57.
13. Kolekar MH, Kumar V. Automated classification of focal and
non-focal EEG signals based on bivariate empirical mode
decomposition. Biomed Signal Image Process Patient Care.
2017;1:1-20.
14. Gupta V, Priya T, Kumar A, Pachori RB, Rajendra Acharya U.
Automated detection of focal EEG signals using features
extracted from flexible analytic wavelet transform. Pattern
Recogn Lett. 2017;94:180-188.
15. Truong ND, Nguyen AD, Kuhlmann L, Bonyadi MR. Integer
convolutional neural network for seizure detection. IEEE J
Emerg Select Topic Circuits Syst. 2018;8(4):849-857.
16. Pushpendrasingh PRB. Classification of focal and nonfocal
EEG signals using features derived from fourier-based rhythms.
J Mech Med Biol. 2017;17(07):1-16.
17. Sharma R, Kumar M, Pachori RB, Acharya UR. Decision sup-
port system for focal EEG signals using tunable-Q wavelet
transform. J Comput Sci. 2017;20:52-60.
18. Bhattacharyya A, Pachori RB, Acharya UR. Tunable-Q wavelet
transform based multivariate sub-band Fuzzy entropy with appli-
cation to focal EEG signal analysism. Entropy. 2017;19(3):1-14.
19. Sharma R, Kumar M, Abhinav PRB, Acharya UR. An auto-
matic detection of focal EEG signals using new class of
SRINATH AND GAYATHRI 11
time–frequency localized orthogonal wavelet filter banks.
Knowl-Based Syst. 2017;118:217-227.
20. Dalal M, Tanveer M, Ram P. Automated identification system
for focal EEG signals using fractal dimension of FAWT based
sub-bands signals. IEEE Int Conf Mach Intell Signal Process.
2017;748:22-24.
21. Deivasigamani S, Senthilpari C, Yong WH. Classification of
focal and nonfocal EEG signals using ANFIS classifier for
epilepsy detection. Int J Imaging Syst Technol. 2016;26(4):
277-283.
22. Abhinaya B, Charanya D, PalaniThanaraj K. Feature extraction
and selection of a combination of entropy features for real-time
epilepsy detection. Intl J Eng Comput Sci. 2016;5(4):16073-
16078.
23. Sharma R, Pradip S, Pachori RB. Automated focal EEG signal
detection based on third order cumulant function. Biomed Sig-
nal Process Control. 2020;58:101856.
24. Sharma R, Pachori RB, Acharya UR. An integrated index for
the identification of focal electroencephalogram signals using
discrete wavelet transform and entropy measures. Entropy.
2015;17:5218-5240.
25. Sharma R, Pachori RB, Acharya UR. Application of entropy
measures on intrinsic mode functions for the automated identi-
fication of focal electroencephalogram signals. Entropy. 2015;
17:669-691.
26. Gupta V, Nishad A, Pachori RB. Focal EEG signal detection
based on constant-bandwidth TQWT filter-banks. IEEE Int
Conf Bioinf Biomed. 2018;1:2597-2604.
27. Bern-Barcelona EEG Dataset. Department of Neurology, University
of Bern, Switzerland. http://ntsa.upf.edu/downloads/andrzejak-
rg-schindler-k-rummel-c-2012-nonrandomness-nonlinear-
dependence-andmple.com. 2012. Accessed May 17, 2019.
How to cite this article: Srinath R, Gayathri R.
Detection and classification of
electroencephalogram signals for epilepsy disease
using machine learning methods. Int J Imaging
Syst Technol. 2020;1–12. https://doi.org/10.1002/
ima.22486
12 SRINATH AND GAYATHRI
A preview of this full-text is provided by Wiley.
Content available from International Journal of Imaging Systems and Technology
This content is subject to copyright. Terms and conditions apply.