Wavelet-based feature extraction for classification of epileptic seizure EEG signal

Abstract

Electroencephalogram (EEG) signal-processing techniques are the prominent role in the detection and prediction of epileptic seizures. The detection of epileptic activity is cumbersome and needs a detailed analysis of the EEG data. Therefore, an efficient method for classifying EEG data is required. In this work, a constructive pattern recognition strategy for analysing EEG data as normal and epileptic seizure has been proposed. With this strategy, the signals were decomposed into frequency sub-bands using discrete wavelet transform (DWT). principal component analysis (PCA) and linear discriminant analysis (LDA) are applied to reduce the dimensionality of EEG data. These reduced features were used as input to Naïve Bayes and K-Nearest Neighbour Classifier to classify normal or epileptic seizure signal. The performance of classifier was evaluated in terms of accuracy, sensitivity and specificity. The experimental results show that PCA with Naïve Bayes classifier provides 98.6% accuracy and LDA with Naïve Bayes classifier attains improved result of 99.8% accuracy. Also, the result shows that PCA, LDA with K-NN achieves 98.5% and 100% accuracy. This evaluation is used to propose a reliable, practical epilepsy detection method to enhance the patient’s care and quality of life.

Full text

Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=ijmt20
Download by: [Vellore Institute of Technology], [Sharmila ashok] Date: 15 November 2017, At: 21:08
Journal of Medical Engineering & Technology
ISSN: 0309-1902 (Print) 1464-522X (Online) Journal homepage: http://www.tandfonline.com/loi/ijmt20
Wavelet-based feature extraction for classification
of epileptic seizure EEG signal
A. Sharmila & P. Mahalakshmi
To cite this article: A. Sharmila & P. Mahalakshmi (2017) Wavelet-based feature extraction for
classification of epileptic seizure EEG signal, Journal of Medical Engineering & Technology, 41:8,
670-680, DOI: 10.1080/03091902.2017.1394388
To link to this article: http://dx.doi.org/10.1080/03091902.2017.1394388
Published online: 09 Nov 2017.
Submit your article to this journal
Article views: 3
View related articles
View Crossmark data

RESEARCH ARTICLE
Wavelet-based feature extraction for classification of epileptic seizure
EEG signal
A. Sharmila and P. Mahalakshmi
School of Electrical Engineering, VIT University, Vellore, Tamilnadu, India
ABSTRACT
Electroencephalogram (EEG) signal-processing techniques are the prominent role in the detec-
tion and prediction of epileptic seizures. The detection of epileptic activity is cumbersome and
needs a detailed analysis of the EEG data. Therefore, an efficient method for classifying EEG data
is required. In this work, a constructive pattern recognition strategy for analysing EEG data as
normal and epileptic seizure has been proposed. With this strategy, the signals were decom-
posed into frequency sub-bands using discrete wavelet transform (DWT). principal component
analysis (PCA) and linear discriminant analysis (LDA) are applied to reduce the dimensionality of
EEG data. These reduced features were used as input to Naïve Bayes and K-Nearest Neighbour
Classifier to classify normal or epileptic seizure signal. The performance of classifier was eval-
uated in terms of accuracy, sensitivity and specificity. The experimental results show that PCA
with Naïve Bayes classifier provides 98.6% accuracy and LDA with Naïve Bayes classifier attains
improved result of 99.8% accuracy. Also, the result shows that PCA, LDA with K-NN achieves
98.5% and 100% accuracy. This evaluation is used to propose a reliable, practical epilepsy detec-
tion method to enhance the patient’s care and quality of life.
ARTICLE HISTORY
Received 6 April 2017
Revised 4 October 2017
Accepted 16 October 2017
Published online 9 November
2017
KEYWORDS
Epileptic; discrete wavelet
transform; PCA; LDA; Naïve
Bayes and K-NN classifier
1. Introduction
Epilepsy can be potentially life threatening with brain
failure, heart and lung failure, head trauma due to
accidents and sudden unexpected death. Even under-
stated epileptic can cause insignificant harm in the
brain. Long-term problems such as fall in intelligence
quotient (IQ), depression, suicide, Social problems may
lead to reduce the quality of life. So the diagnosis of
epilepsy is the most important in the existing develop-
ment of research. The main challenge of this work is
to detect the epilepsy seizure in order to maintain
independency in patient’s life and also to help person
with epilepsy to lead full and productive life.
Epilepsy is one of the common neurological dis-
order that affects the nervous system and is character-
ised by the transient and unexpected electrical
disturbance of the brain. All brain functions including
feeling, seeing, thinking and moving muscles depend
on electrical signals passed between nerve cells in the
brain. Epilepsy is also known as a seizure disorder. A
seizure occurs when too many nerve cells in the brain
“fire”too quickly causing an “electrical storm”.
According to the WHO [1], there are about 50 million
people worldwide who are suffering from epilepsy. In
about 70% of people with epilepsy, the cause is not
known and in 30% of the people most common
causes are due to head trauma, infection of brain tis-
sue, brain tumour and stroke, heredity and prenatal
disturbance of brain development.
The electroencephalogram (EEG) is the most influ-
ential technique in the detection of epileptic seizures.
EEG records the electric activity of the cerebral cortex
in the brain and it is used to measure or diagnose
brain diseases. EEG monitoring is an important factor
when identifying conditions, particularly of patients
with epilepsy. The recording data of the human EEGs
are carried out by placing the electrodes on the scalp,
voltage range of the scalp EEG lie between 10 and
100 lV. The main concern in EEG analysis is how to
differentiate seizure from normal EEG signal.
Conventionally, seizure activities are visually
inspected from EEG signals by trained physicians. But
this is a time-consuming and cumbersome in case of
long EEG recordings. So, a rapid detection and classifi-
cation of seizure activity would give a great support to
quantitative analysis and interpretation. In modern
years, different methods have been developed to auto-
matically classify the normal or abnormal EEG signals
without spending hours for visual inspection.
CONTACT P. Mahalakshmi pmahalakshmi@vit.ac.in School of Electrical Engineering, VIT University, Vellore, Tamilnadu, India
ß2017 Informa UK Limited, trading as Taylor & Francis Group
JOURNAL OF MEDICAL ENGINEERING & TECHNOLOGY, 2017
VOL. 41, NO. 8, 670–680
https://doi.org/10.1080/03091902.2017.1394388
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

The process of automatic classification is achieved
using pattern recognition. The various stages in pat-
tern recognition are feature extraction, feature reduc-
tion and classification as shown in Figure 1.
Feature extraction is a technique for extracting the
hidden characteristics of the signal such as amplitude,
frequency, mean value, etc., to interpret the signal.
Since the numbers of features are larger due to multi-
channel data, the computation time and memory
requirement will be more. Therefore, the number of
the extracted features can be reduced using a tech-
nique called feature reduction. The reduced features
are fed as input to the classifier. In initial stage, fea-
tures are extracted from raw EEG data using time
domain, frequency domain or time–frequency domain
methods. Meanwhile, the EEG signals are non-station-
ary in nature, it is best suitable to use time–frequency
domain like the wavelet transform for feature extrac-
tion. It offers both time and frequency information of
a signal, and thus, it is possible to get and confine fea-
tures in the data such as the seizure EEG activity
accurately. In the second stage, the features are
reduced to obtain lower dimensionality of the data,
and in last stage, reduced features are used as the
input for the classifier to differentiate normal and epi-
leptic seizure. However, the classification performance
is mostly dependent on the features that are being
used to characterise the original EEG. Therefore, fea-
ture extraction is one of the most significant constitu-
ents of pattern recognition. It suggestively contributes
to the performance of the classifier and reduces data
size without losing its characteristic power. Thus, it is
important to propose an effective feature extraction
method instead of designing the structure of a com-
plex classifier.
In this work, publicly available data provided by
Department of Epitology at Bonn University, Germany.
Discrete Wavelet Transform (DWT) has been applied
for the time-frequency analysis of EEG signals for the
classification using wavelet coefficients. EEG signals
were decomposed into frequency sub-bands using
DWT, and then, a set of statistical features was
extracted from these sub-bands to represent the distri-
bution of wavelet coefficients. PCA and LDA are used
to reduce the dimension of the data. Then, these fea-
tures were used as an input to the Naïve Bayes and k-
NN classifier to distinguish outputs: normal or seizure.
The accuracy of the various classifiers will be evaluated
and cross-compared, advantages and limitations of
each technique will be discussed. The simulation
shows that k-NN with PCA and LDA can always per-
form better than Naïve Bayes with PCA and LDA. Also,
among PCA and LDA, the best performance is attained
in using LDA.
This paper is organised as follows: Section 2 dis-
cusses other studies performed by using the same epi-
leptic EEG dataset. In Section 3, the clinical data used
in this study are presented, followed by feature extrac-
tion method based on the DWT and dimensional
reduction using PCA and LDA in the subsections. In
Section 4, we have analysed and compared the exist-
ing and proposed schemes in terms of device discov-
ery times. Section 5 concludes this paper.
2. Literature review
Kalayci et al. used wavelet transform to capture some
specific characteristic features of the EEG signals and
then combined with ANN to get satisfying classifica-
tion result [1]. Nigam et al. [2] described a method for
automated detection of epileptic seizures from EEG
signals using a multistage non-linear pre-processing fil-
ter for extracting two features: relative spike amplitude
and spike occurrence frequency. Then, they fed those
features to a diagnostic artificial neural network [2].
In the work of Jahankhani et al., the EEGs were
decomposed with wavelet transform into different
sub-bands and some statistical information was
extracted from the wavelet coefficients. Radial basis
function network (RBF) and multi-layer perceptron net-
work (MLP) were utilised as classifiers [3]. Subasi
decomposed the EEG signals into time–frequency
Figure 1. Pattern Recognition of EEG Signal to interpret abnormality.
JOURNAL OF MEDICAL ENGINEERING & TECHNOLOGY 671
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

representations using discrete wavelet transform.
Some features based on DWT were obtained and
applied for different classifiers for epileptic EEG classifi-
cation, such as feed-forward error back-propagation
artificial neural network (FEBANN), dynamic wavelet
network (DWN), dynamic fuzzy neural network (DFNN)
and mixture of expert system (ME) [4–6]. Ubeyli
employed wavelet analysis with combined neural net-
work model to discriminate EEG signals. The EEGs
were decomposed into time–frequency representations
using DWT, and then, statistical features were calcu-
lated. Then, a two-level neural network model was
used to classify three types of EEG signals. The results
proved that utilising combined neural network model
achieved better classification performance than the
stand-alone neural network model [7].
Ocak detected epileptic seizures based on approxi-
mate entropy (ApEn) and discrete wavelet transform.
EEG signals were firstly decomposed into approxima-
tion and detail coefficients using DWT, and then, ApEn
values for each set of coefficients were computed.
Finally, surrogate data analysis was used on the ApEn
values to classify EEGs [8].
Guo et al. proposed epileptic seizure detection
using multiwavelet transform-based approximate
entropy and artificial neural networks [9]. Orhan et al.
studied EEG signals classification using k-means clus-
tering and a multilayer perceptron neural network
model [10]. Iscan et al. worked on classification of elec-
troencephalogram signals with combined time and fre-
quency features [11]. Wang et al. investigated the best
basis-based wavelet packet entropy feature extraction
and hierarchical EEG classification for epileptic detec-
tion [12]. Xie et al. proposed a wavelet-based sparse
functional linear model with applications to EEG seiz-
ure detection and epilepsy diagnosis [13]. Janjarasjitt
studied on the classification of the epileptic EEGs by
using the wavelet based scale variance feature [14].
Yatindra kumar et al. aimed to present detection of
epileptic seizures from EEG data recorded for normal
subjects and epileptic patients using DWT-based ApEn
and artificial neural network [15].
3. Clinical EEG data and method
3.1. Clinical EEG data set
The EEG data used in this study were obtained from
Department of Epitology at University of Bonn,
Germany [34]. The data consist of five sets that are
called set A to E and each data set takes 23.6 s consist-
ing of 100 EEG segments recorded on the scalp by sin-
gle channel. The set A and set B include surface EEG
recordings that are collected from five healthy subjects
while eyes were opened and closed, respectively using
standardised electrode placement scheme. The dataset
of set C, D and E were obtained from five epileptic
patients undergoing pre-surgical evaluations. Set C
was recorded on patients before epileptic attack at
hemisphere hippocampal formation and set D was
recorded from epileptogenic zone. The data set E were
recorded from patients during an epilepsy occurrence
using depth electrodes placed within epileptogenic
zone. The data were recorded with 128-channel ampli-
fier system and digitised through 12-bit A/D converter
with a sampling frequency of 173.61 Hz. The data from
set A to E consists of 100 files, and each file contains
4097 successive EEG signals. The 4097 successive EEG
signals were divided into eight sets of 512 signals in
total, and last EEG signal were deleted. In this work,
only data from set A and set E have been taken, so a
total of 1600 sets were obtained for these sets. The
performance can be evaluated from the results
obtained for 800 training set and 800 testing set and
also for 600 training set and 1000 testing set for nor-
mal and epileptic patients.
3.2. Statistical feature extraction using discrete
wavelet transform (DWT)
Wavelet transforms (WT) are widely applied in biomed-
ical engineering areas for solving a variety of real-life
problems. WT provides a more flexible way of time-fre-
quency representation of a signal by allowing the use
of variable-sized windows. WT has gained practical
interest in extracting the valuable information
embedded on the EEG signals due to its ability in cap-
turing valuable frequency information at low-frequency
bands and time information at high-frequency bands
[16–19]. EEG signals are non-stationary in nature, and
they contain high-frequency information with short
time period and contain low-frequency information
with long time period [20]. Therefore, by analysing the
biomedical signals at different time and frequency reso-
lutions, WT is able to pre-process the biomedical signals
efficiently in the feature extraction stage.
The continuous wavelet transform (CWT) [21]ofa
signal, xt
ðÞ
;is the integral of the signal multiplied by
scale and shifted versions of a wavelet function wand
is defined by,
CWTða;bÞ¼ð1
1
xðtÞ1
ffiffiffiffiffiffi
jaj
pWtb
a

dt (1)
where aand bare called the scaling and shifting
parameters, respectively. Calculating wavelet
672 A. SHARMILA AND P. MAHALAKSHMI
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

coefficients at every possible scale is computationally
a very expensive task. Instead, if the scales and shifts
are selected based on powers of two, so-called dyadic
scales and positions, then the wavelet analysis will be
much more efficient. Such analysis is obtained from
DWT which is defined as,
DWT j;kðÞ¼1
ffiffiffiffiffiffiffi
j2jj
pð1
1
xt
ðÞwt2jk
2j

dt (2)
where aand bare replaced by 2jand k2j, respectively.
In the first step of the DWT, the signal is simultan-
eously passed through a g[n] as a LP and h[n]asaHP
filters with the cut-off frequency being the one fourth
of the sampling frequency. The use of a group of fil-
ters to divide up a signal into various spectral compo-
nents is termed sub-band coding. This procedure is
known as multiresolution decomposition of a signal
x[n]. Each stage of this scheme consists of two digital
filters and two down-samplers by 2. The first filter, h[.]
is the discrete mother wavelet, high-pass in nature,
and the second, g[.] is the mirror version, low pass in
nature. The down-sampled outputs of first high-pass
and low-pass filters provide the detail, D1 and approxi-
mation, A1, respectively [22–24]. At each step of
decomposition process, the frequency resolution is
doubled through filtering and the time resolution is
halved through down sampling.
Figure 2 represents the fifth-level wavelet decom-
position of a signal. In this representation, the coeffi-
cients A1, D1, A2, D2, A3, D3, A4, D4, A5 and D5
represent the frequency content of the original signal
within the bands 0-fs/4, fs/4-fs/2, 0-fs/8, fs/8-fs/4, 0-fs/
16, fs/16-fs/8, 0-fs/32, fs/32-fs/16, 0-fs/64 and fs/64-fs/
32, respectively, where fs is the sampling frequency of
the original signal x[n].
Selection of appropriate wavelet and the number of
levels of decomposition is very important in analysis of
signals using DWT. The number of levels of decompos-
ition is chosen based on the dominant frequency com-
ponents of the signal. The levels are chosen such that
those parts of the signal that correlate well with the
frequencies required for the classification of the signal
are retained in the wavelet coefficients. Since the EEG
signals do not have any useful frequency components
above 30 Hz, the number of levels was chosen to be 5.
Thus, the signal is decomposed into the details D1–D5
and one final approximation, A5. These approximation
and detail records are reconstructed from the
Daubechies 4 (DB4) wavelet filter [22]. The extracted
wavelet coefficients provide a compact representation
that shows the energy distribution of the EEG signal in
time and frequency. In order to further decrease the
dimensionality of the extracted feature vectors, statis-
tics over the set of the wavelet coefficients was used
[25]. The following statistical features were used to
represent the time–frequency distribution of the EEG
signals:
1. Mean of the absolute values of the coefficients in
each sub-band.
2. Average power of the wavelet coefficients in each
sub-band.
3. Standard deviation of the coefficients in each sub-
band.
3.3. Feature reduction methods
3.3.1. Principal component analysis (PCA)
PCA is a useful statistical technique for finding pat-
terns in data of high dimension and it is unsupervised
approach. It is a way of identifying patterns in data,
and expressing the data in such a way as to focus
their similarities and differences. Since patterns in data
can be hard to find in data of high dimension, where
the amenity of graphical representation is not avail-
able, PCA is a powerful tool for analysing data [26].
PCA extracts a lower dimensional space by analysing
the covariance structure of multivariate statistical
observations. The objective of PCA is to represent data
in a space that best expresses the variation in a sum-
squared error sense. Mostly useful for segmenting
signals from multiple sources. Its properties are as
Figure 2. Level-five wavelet decomposition of input EEG signal.
JOURNAL OF MEDICAL ENGINEERING & TECHNOLOGY 673
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

follows: (i) it maximises the variance of the extracted
features; (ii) the extracted features are uncorrelated;
(iii) it finds the best linear approximation; (iv) it maxi-
mises the information contained in the extracted fea-
tures. The main advantage of PCA is that to find
patterns in the data and compress the data by reduc-
ing the number of dimensions without losing much
information. The idea to use PCA is due to the fact, in
most of EEG data; there is a large amount of redun-
dant information unnecessary for diagnostic applica-
tions. The computational steps are as follows:
i. Calculate the covariance matrix Sfrom the input
data.
ii. Compute the eigenvalues and eigenvectors of S
and sort them in a descending order with respect
to the eigenvalues.
w eig decompositionðS¼Rn
ði¼1ÞðximÞðximÞTÞ
(3)
where “n”is the number of instances, xiis the i-th
instance, and “m”is mean vector of the input data.
iii. Form the actual transition matrix by taking the
predefined number of components (eigenvectors).
iv. Finally, multiply the original feature space with
the obtained transition matrix, which yields a
lower-dimensional representation.
It can be shown that this representation of
Equation (3) minimises a squared error criterion
[27,28].
3.3.2. Linear discriminant analysis (LDA)
LDA is a supervised approach and it works by creat-
ing a new variable that is a combination of the ori-
ginal predictors. This is done in such a way that the
differences between the predefined groups, with
respect to the new variable, are maximised. The aim
is to combine (weight) the predictor scores in some
way so that a single new composite variable, the dis-
criminant score, is produced. It is possible to view
this as an extreme data dimension reduction tech-
nique that compresses the p-dimensional predictors
into a one-dimensional line. At the end of the pro-
cess, it is hoped that each class will have a normal
distribution of discriminant scores but with the larg-
est possible difference in mean scores for the classes.
Indeed, the degree of overlap between the discrimin-
ant score distributions can be used as a measure of
the success of the technique. Discriminant scores are
calculated by a discriminant function which has
the form:
D¼w1Z1þw2Z2þw3Z3þ:::::::::::::::: þwpZp(4)
Discriminant score is a weighted linear combination
of the predictors. The weights are estimated to
maximise the differences between class mean dis-
criminant scores.
Generally, those predictors which have large dissim-
ilarities between class means will have larger
weights, at the same time weights will be small
when class means are similar [30].
3.4. Naive Bayes classifier
A naive Bayes classifier is a probabilistic classifier that
makes use of Bayesian theory and is the optimal
supervised learning method if the predictors are inde-
pendent (uncorrelated), given the class. This means
that, if its assumptions are met, it is guaranteed to
produce the most accurate predictions. It does, there-
fore, place an upper limit on what a classifier can
achieve. Despite the obvious over-simplification of
these assumptions, the naive Bayes classifier has
proved be a very effective supervised learning algo-
rithm on real data, where the assumptions do not
apply. Bayesian methods are intentionally simplistic
manner to obtain class predictions. The naivety arises
from the assumptions about the independence of the
predictors, which are unlikely to be valid in real data.
Consequently, the data in a naive Bayes classifier will
come from unknown, but estimated, probability distri-
butions that share some interdependency. In reality, it
is not even necessary to use Bayesian methods to pro-
duce these predictions because the initial Bayesian
model can be rewritten in a form that is computation-
ally more feasible using maximum likelihood techni-
ques. Despite these apparently important deviations
from the underlying assumptions, the technique
appears to work well. This is because the predictions
will be accurate as long as the probability of being in
the correct class is greater than that for any of the
incorrect classes. In other words, only an approximate
solution is required that has the correct rank order for
class membership probabilities.
The Bayes rule assigns a case to a class (k) which
has the largest posterior probability given the linear
combination of predictors x
ðÞ
:classi¼maxkpkjx
ðÞ

.
Bayesian methods are then used to estimate pkjx
ðÞ
,so
that a class identity can be assigned [29].
Let “C
1
”and “C
2
”be two classes (Normal and
Epileptic seizure respectively), “N”be the total number
of training samples and “N1”,“N2”belong to “C
1
”and
674 A. SHARMILA AND P. MAHALAKSHMI
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

“C
2
”, respectively. Then, the probabilities P(C
1
)¼N1/N
and P(C
2
)¼N2/N. These probabilities are called “prior
probabilities”.P(X/C
i
) is referred as likelihood function
of “C
i
”with respect to sample “X”. According to Bayes
rule,
PC
i=X
ðÞ
¼X=Ci
ðÞ
PC
i
ðÞ
=PX
ðÞ (5)
where P(C
i
/X) is the probability that sample “X”
belongs to class “C
i
”and P(X) is the probability distri-
bution function of “X”.
PðXÞ¼X2
i¼1PðX=CiÞPðCiÞ(6)
Bayes classification rule can be stated as If P(C
1
/
X)>P(C
2
/X), “X”belongs to Class “C
1
”and If P(C
2
/
X)>P(C
1
/X), “X”belongs to Class “C
2
”.
3.5. k-NN classifier
k-Nearest neighbour (k-NN) classifier [30] is a non-para-
metric, non-linear and relatively simple classifier. It
classifies a new sample on the basis of measuring the
“distance”to a number of patterns that are kept in
memory. The class that k-NN classifier determines for
this new sample is decided by the pattern thatmost
resembles it, that is, the one that has the smallest dis-
tance to it. The common distance function used in the
KNN classifier is the Euclidean distance. Instead of tak-
ing the single nearest sample, it is normally taking a
majority vote from the k-nearest neighbours [31,32].
The parameter khas to be selected in practice. In this
work, kis chosen to 2.
To classify an unknown data, the computational
steps are as follows: (i) Compute distance to other
training data; (ii) Identify knearest neighbours; (iii) Use
class labels of nearest neighbours to determine the
class label of unknown data.
Compute distance between two points, that is,
Euclidean distance:
dðp;qÞ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
i
ðpiqiÞ2
r(7)
From Euclidean distance, the class from nearest
neighbour list has been determined. The majority vote
of class labels among the k-nearest neighbours has
been considered. According to distance, vote can be
weighted using weight factor (w)¼1/d
2
. The value of
khas to be chosen. If kis too small, sensitive to noise
points. If kis too large, neighbourhood may include
points from other classes. The parameter khas to be
selected in practice. In this work, kis chosen to 2.
3.6. Statistical parameters
The performance of classifier is evaluated by using
three constraints that is sensitivity, specificity and clas-
sification accuracy. The description of these constraints
is as follows:
Sensitivity: The number of correctly detected positive
patterns/total number of actual positive patterns.
A positive pattern indicates a detected seizure.
Specificity: The number of correctly detected negative
patterns/total number of actual negative patterns.
A negative pattern indicates a detected normal.
Classification accuracy: The number of correctly clas-
sified patterns/total number of patterns.
4. Results and discussion
The purpose of this study was to investigate the wave-
let-based approach for obtaining representative fea-
tures from EEG data. For this purpose, an EEG
classification system, which used the proposed wave-
let-based feature extraction method, was created. In
this study, EEG signals of normal and epileptic patients
in order to perform a comparison between PCA and
LDA by using SVM were used. EEG recordings were
divided into sub-band frequencies such as a,b,dand
hby using DWT. The wavelet decomposition of sample
EEG epochs, the approximation and detail coefficients
of data set A and E are shown in Figures 3 and 4.
All EEG epochs from both normal subjects and epi-
leptic patients are decomposed into sub-bands D1, D2,
D3, D4, D5 and A5 and their frequencies correspond-
ing to different levels of decomposition for
Daubechies four filter wavelet with a sampling fre-
quency of 173.6 Hz are shown in Table 1.
After normalisation, the EEG signals were decom-
posed using wavelet transform and statistical features
were extracted from wavelet sub-band frequencies.
Then dimensions of these features are reduced by using
PCA and LDA. A classification system based on k-NN and
naïve Bayes were implemented using these data as
inputs.
The objective of the modelling phase in this work
was to develop classifiers that are able to identify any
input combination as belonging to either one of the
two classes: normal or epileptic. Experimental groups
A and E consist of 100 files, and each file contains
4097 successive EEG signals. These signals were div-
ided into eight sets of 512 signals in total, and the last
EEG signal was deleted. Thus, 800 sets of 512 signals
were created from 100 files and 1600 sets were cre-
ated from set A and set E. For developing classifiers,
800 sets were randomly taken from the 1600 sets and
JOURNAL OF MEDICAL ENGINEERING & TECHNOLOGY 675
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

used for training the classifier, and remaining 800 sets
were used for testing the developed models. The class
distribution, number of training and test data set are
summarised in Tables 2 and 3.
The problem involves classification into two classes:
normal or epileptic. The performance measure used
were sensitivity, specificity and accuracy. Sensitivity,
also called the true positive ratio, is calculated by the
formula:
Sensitivity TPR
ðÞ
¼TP
TP þFN 100%
Specificity, also called the true negative ratio, is cal-
culated by the formula:
Specificity TNR
ðÞ
¼TN
TN þFP 100%
Accuracy is calculated by the formula:
Accuracy ¼TN þTP
TN þTP þFN þFP 100%
Epileptic seizure detection in EEG can be under-
stood as a kind of pattern recognition theory. It
consists of data acquisition, signal processing, feature
extraction, feature reduction and seizure detection. A
novel EEG signal classification is proposed, which is
based on DWT, the dimension reduction based on
PCA and LDA and classification based on k-NN and
Naïve Bayes. The procedure of the proposed system
can be summarised as follows:
i. The features calculated with statistical features
parameter from time–frequency domain using
DWT.
ii. Feature reduction based on PCA and LDA algo-
rithm was applied to extracted features to reduce
the dimensionality. This step is implemented to
get rid of irrelevant features that are redundant
and even degrade the performance of classifier.
iii. The classification process for epileptic seizure
detection is carried out using k-NN and Naïve
Bayes classification.
The procedure was repeated on EEG recordings of
healthy subject and epileptic patients. In this work, the
Figure 3. Wavelet decomposition of sample EEG epoch of set A.
676 A. SHARMILA AND P. MAHALAKSHMI
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

training process was carried out using PCA þk-NN,
LDA þk-NN, PCA þNaïve Bayes (NB) and LDA þnaïve
Bayes (NB). After training, we used two different fea-
ture reduction methods and get the test results that
are shown in Tables 4 and 5.
As seen in Tables 4 and 5, the classification rate
with LDA feature reduction is highest than PCA. Also,
simulation shows that k-NN classification using PCA,
Figure 4. Wavelet decomposition of sample EEG epoch of set E.
Table 1. Wavelet decomposition and their
frequencies.
Decomposed signal Frequency (Hz)
D1 43.4
D2 21.7–43.4
D3 10.8–21.7
D4 5.4–10.8
D5 2.7–5.4
A50–2.7
Table 2. Number of training and test sets.
Class Training set Test set Total
Epileptic 400 400 800
Normal 400 400 800
Total 800 800 1600
Table 3. Number of training and test sets.
Class Training set Test set Total
Epileptic 300 500 800
Normal 300 500 800
Total 600 1000 1600
Table 4. Number of training and test sets (as in Table 2).
Technique Accuracy (%) Sensitivity (%) Specificity (%)
Computational Time (sec)
PCA & NB 97 94 100 0.987
LDA & NB 99.75 99.5 100 1.057
PCA & k-NN 98.5 97.08 100 0.895
LDA & k-NN 100 100 97.2 0.924
Table 5. Number of training and test sets (as in Table 3).
Technique Accuracy (%) Sensitivity (%) Specificity (%)
Computational time (sec)
PCA & NB 98.6 100 97.2 1.509
LDA & NB 99.8 99.6 100 0.992
PCA & k-NN 98.5 97.08 100 1.295
LDA & k-NN 100 100 100 0.933
JOURNAL OF MEDICAL ENGINEERING & TECHNOLOGY 677
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

LDA can perform better than naïve Bayes and gives an
accuracy of 98.5% and 100% for equal and unequal
training and testing set. The classification accuracy of
LDA with k-NN is 100%, which is highest of all techni-
ques for equal and unequal training and testing set.
According to computational time, the classification pro-
cess using LDA with naïve Bayes classifier and k-NN clas-
sifier is relatively longer than PCA with naïve Bayes and
k-NN classifier for data set as shown in Table 2. For the
data set as shown in Table 3, the computational time
for PCA with naïve Bayes and k-NN classifier is longer
than LDA with naïve Bayes and k-NN classifier.
Table 6 presents the comparison between results
obtained from proposed method and existing
methods. Only, the methods that used the same data
set are included for the purpose of comparison. The
accuracy obtained from the proposed method gives
the best result with LDA feature reduction technique
and k-NN classifier (100%). It can be seen that our pro-
posed method is fast enough for real-time EEG seizure
detection. The programme was written in MATLAB
R2014a and personal computer is HP 248 G1
Notebook with a 1.6 GHz CPU and 8GB of memory.
Conclusions
EEG recordings have turn out to be a widespread
means for seizure detection and diagnosis. Visually,
Table 6. Shows comparison of automated epileptic seizure detection methods using the open source dataset provided by
University of Bonn [33].
Authors Year Feature extraction Feature reduction Feature classification Accuracy (%)
Guler & Ubeyli [34] 2005 Statistical features from wavelet
transform
None Adaptive neuro fuzzy
system
98.68
Sadati et al. [35] 2006 Statistical features from wavelet
transform
None Adaptive neuro fuzzy
system
85.6
Guler & Ubeyli [36] 2007 Wavelet transform and
Lyapunov exponents
None Support vector machine
and probabilistic
neural network (PNN)
99.28
Subasi [6] 2007 Statistical features from wavelet
transform
None Mixture of expert model 95
Polat & Gunes [37] 2008 Wavelet transform, fast Fourier
transform and autoregressive
model
Distance-based data
reduction
Decision tree 99.32
Ocak [38] 2008 Wavelet transform, approximate
entropy and genetic
algorithm
None Learning vector
quantisation
98
Mousavi et al. [39] 2008 Wavelet transform & auto
regressive model
None Artificial neural
network(ANN)
96
Ubeyli [7] 2008 Statistical features from wavelet
transform
None Mixture of expert model 93.17
Ocak [8] 2009 Wavelet transform and approxi-
mate entropy
None Surrogate data analysis 96.65
Guo et al. [46] 2009 Wavelet transform & relative
wavelet energy
None ANN 95.2
Ubeyli [18] 2009 Statistical features from wavelet
transform
None Probabilistic neural
network
97.63
Guo et al. [9] 2010 Wavelet transform and approxi-
mate entropy
None ANN 99.85
Subasi & Gursoy [40] 2010 Statistical features from wavelet
transform
PCA, ICA and LDA SVM 100
Lim et al. [41] 2010 Statistical features from wavelet
transform
None LS-SVM 100
Wang et al. [12] 2011 Wavelet transform and
Shannon entropy
None k-NN classifier 100
Orhan et al. [10] 2011 Statistical features from wavelet
transform
None k-NN classifier, ANN 100
100
Sezer et al. [42] 2012 Statistical features from wavelet
transform
PCA ANN 100
Song & Zhang [43] 2013 Wavelet transform and permu-
tation entropy, sample
entropy and Hurst exponent
Genetic algorithm
(feature selector)
Extreme learning
machine (ELM)
94.8
Guangyi Chen [44] 2014 Dual tree complex wavelet
transform
None k-NN 100
Yatindra Kumar et al. [15] 2014 Wavelet transform and approxi-
mate entropy (ApEn)
None Artificial neural network,
SVM
100
Gajic et al. [45] 2015 Statistical features from wavelet
transform
Scatter matrices Quadratic 99
Present reporting Statistical features from wavelet
transform
PCA
PCA
LDA
LDA
Naïve Bayes
k-NN
Naïve Bayes
k-NN
98.6
98.5
99.8
100
678 A. SHARMILA AND P. MAHALAKSHMI
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

epileptic seizure detection by trained clinician from
the longtime EEG recordings is a very time-consuming
and expensive process. Therefore, it is desirable to
choose a reliable, simple and fast method for feature
extraction and classification from EEG signals. In this
paper, a new method for detecting epilepsy seizures
have been introduced by using statistical features
obtained from DWT and feature reduction techniques
such as PCA and LDA for classifying the EEG signal as
normal or epileptic seizure with k-NN and Naïve Bayes
classifier. The achievement of the proposed technique
is confirmed through comparing the performance of
classification problems from other researchers. This
method achieves the highest classification accuracy of
100% using LDA as feature reduction technique and
k-NN classifier for the database developed by the
University of Bonn. This proposed method is fast and
easy to implement, so it should be a practical method
for the detection and treatment of epileptic seizure
disorder patients.
Disclosure statement
No potential conflict of interest was reported by the authors.
References
[1] Kalayci T, Ozdamar O. Wavelet preprocessing for auto-
mated neural network detection of EEG spikes. IEEE
Eng Med Biol Mag. 1995;14:160–166.
[2] Nigam V, Graupe D. A neural-network-based detection
of epilepsy. Neurol Res. 2004;26:55–60.
[3] Jahankhani P, Kodogiannis V, Revett K. EEG signal
classification using wavelet feature extraction and
neural networks. In: IEEE John Vincent Atanasoff 2006
international symposium on modern computing
(JVA’06); 2006. p. 52–57.
[4] Subasi A. Epileptic seizure detection using dynamic
wavelet network. Expert Syst Appl. 2005;29:343–355.
[5] Subasi A. Automatic detection of epileptic seizure
using dynamic fuzzy neural networks. Expert Syst
Appl. 2006;31:320–328.
[6] Subasi A. EEG signal classification using wavelet fea-
ture extraction and a mixture of expert model. Expert
Syst Appl. 2007;32:1084–1093.
[7] Ubeyli E. Combined neural network model employing
wavelet coefficients for EEG signals classification.
Digital Signal Process. 2009;19:297–308.
[8] Ocak H. Automatic detection of epileptic seizures in
EEG using discrete wavelet transforms and approxi-
mate entropy. Expert Syst Appl. 2009;36:2027–2036.
[9] Guo L, Rivero D, Pazos A. Epileptic seizure detection
using multiwavelet transform based approximate
entropy and artificial neural networks. J Neurosci
Methods. 2010;193:156–163.
[10] Orhan U, Hekim M, Ozer M. EEG signals classification
using k-means clustering and a multilayer perceptron
neural network model. Expert Syst Appl. 2011;38:
13475–13481.
[11] Iscan Z, Dokur Z, T, D. Classification of electroenceph-
alogram signals with combined time and frequency
features. Expert Syst Appl. 2011;38:10499–10505.
[12] Wang D, Miao D, Xie C. Best basis-based wavelet
packet entropy feature extraction and hierarchical
EEG classification for epileptic detection. Expert Syst
Appl. 2011;38:14314–14320.
[13] Xie S, Krishnan S. Wavelet-based sparse functional lin-
ear model with applications to EEGs seizure detection
and epilepsy diagnosis. Med Biol Eng Comput.
2013;51:49–60.
[14] Janjarasjitt S. Classification of the epileptic EEGs using
the wavelet based scale variance feature. Int J Appl
Biomed Eng. 2010;3:19–25.
[15] Kumar Y, Dewal ML, Anand RS. Epileptic seizures
detection in EEG using DWT-based ApEn and artificial
neural network. SIViP. 2014;8:1323–1334.
[16] Ebrahimpour R, Babakhani K, Arani SAAA, et al.
Epileptic seizure detection using a neural network
ensemble method and wavelet transform. NNW.
2012;22:291–310.
[17] Guo L, Rivero D, Pazos A. Epileptic seizure detection
using multiwavelet transform based approximate
entropy and artificial neural networks. J Neurosci
Meth. 2010;193:156–163.
[18] Ubeyli ED. Combined neural network model employ-
ing wavelet coefficients for EEG signals classification.
Digit Signal Process. 2009;19:297–308.
[19] Zandi AS, Javidan M, Dumont GA, et al. Automated
real-time epileptic seizure detection in scalp EEG
recordings using an algorithm based on wavelet
packet transform. IEEE Trans Biomed Eng. 2010;57:
1639–1651.
[20] Tzallas AT, Tsipouras MG, Fotiadis DI. Automatic seiz-
ure detection based on time-frequency analysis and
artificial neural networks. 2007;2007:80510.
[21] Mohseni H, Maghsoudi A, Kadbi M, et al. Automatic
detection of epileptic seizure using time–frequency
distributions. In: IET 3rd International Conference on
Advances in Medical, Signal and Information
Processing, MEDSIP 2006. Vol. 14; 2006.
[22] Adeli H, Zhou Z, Dadmehr N. Analysis of EEG records
in an epileptic patient using wavelet transform.
J Neurosci Methods. 2003;123:69–87.
[23] Marchant BP. Time–frequency analysis for biosystem
engineering. Biosyst Eng. 2003;85:261–281.
[24] Semmlow JL. Biosignal and biomedical image process-
ing: MATLAB-based applications. New York: Marcel
Dekker, Inc; 2004.
[25] Kandaswamy A, Kumar CS, Ramanathan RP, et al.
Neural classification of lung sounds using wavelet
coefficients. Comput Biol Med. 2004;34:523–537.
[26] Smith LI. A tutorial on principal components analysis.
Cornell University, USA; 2002.
[27] Cao LJ, Chua KS, Chong WK, et al. A comparison of
PCA, KPCA and ICA for dimensionality reduction in
support vector machine. Neurocomputing. 2003;55:
321–336.
[28] Duda RO, Hart PE, Strok DG. Pattern classification.
2nd ed.; 2001. p. 20–25.
JOURNAL OF MEDICAL ENGINEERING & TECHNOLOGY 679
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017

[29] Fielding AH, Cluster and classification techniques for
the biosciences. Cambridge, UK: Cambridge University
Press; 2007.
[30] Cover T, Hart P. Nearest neighbor pattern classifica-
tion. IEEE Trans Inform Theory. 1967;13:21–27.
[31] Orhan U, Hekim M, Ozer M. EEG signals classification
using the K-means clustering and a multilayer percep-
tron neural network model. Expert Syst Appl.
2011;38:13475–13481.
[32] Guo L, Rivero D, Dorado J, et al. Automatic feature
extraction using genetic programming: an application
to epileptic. EEG Classif. 2011;38:10425–10436.
[33] EEG database from University of Bonn. [cited 16 June
2013]. Available from: http://www.epileptologiebonn.
de/cms/front_content.php?idcat¼193
[34] Guler I, Ubeyli ED. Adaptive neuro-fuzzy inference sys-
tem for classification of EEG signals using wavelet
coefficients. J Neurosci Methods. 2005;148:113–121.
[35] Sadati N, Mohseni HR, Magshoudi A. Epileptic Seizure
Detection Using Neural Fuzzy Networks. In: Proc. IEEE
International Conference on Fuzzy Systems,
Vancouver, Canada; 2006. p. 596–600.
[36] Guler I, Ubeyli ED. Multiclass support vector machines
for EEG-signals classification. IEEE Trans Inform
Technol Biomed. 2007;11:117–126.
[37] Polat K, Gunes S. A novel data reduction method:
Distance based data reduction and its application to
classification of epileptiform EEG signals. Appl
Comput. 2008;200:10–27.
[38] Ocak H. Optimal classification of epileptic seizures in
EEG using wavelet analysis and genetic algorithm. Sig
Process. 2008;88:1858–1867.
[39] Mousavi SR, Niknazar M, Vahdat BV, Epileptic Seizure
Detection using AR Model on EEG Signals. Biomedical
Engineering Conference. CIBEC 2008. Cairo
International, 2008 Dec 18-20; Cairo International, 1-4.
[40] Subasi A, Gursoy I. EEG signal classification using PCA,
ICA, LDA and support vector machines. Expert Syst
Appl. 2010;37:8659–8666.
[41] Lima CA, Coelho AL, Eisencraft M. Tackling EEG signal
classification with least squares support vector
machines: a sensitivity analysis study. Comput Biol
Med. 2010;40:705–714.
[42] Sezer E, Is¸ik H, Saracoglu E. Employment and compari-
son of different artificial neural networks for epilepsy
diagnosis from EEG signals. J Med Syst. 2012;36:
347–362.
[43] Song Y, Zhang J. Automatic recognition of epileptic
EEG patterns via Extreme Learning Machine and mul-
tiresolution feature extraction. Expert Syst Appl.
2013;40:5477–5489.
[44] Chen G. Automatic EEG seizure detection using dual-
tree complex wavelet-Fourier features. Expert Syst
Appl. 2014;41:2391–2394.
[45] Gajic D, Djurovic Z, Gligorijevic J, et al. Detection of
epileptiform activity in EEG signals based on time-
frequency and non-linear analysis. Front Comput
Neurosci. 2015;9:1–16.
[46] Guo L, Rivero D, Seoane J, Pazos A. Classification of
EEG signals using relative wavelet energy and artificial
neural networks. In: Proc. ACM/SIGEVO, Summiton
Genetic and Evolutionary Computation, New York,
USA; 2009. p. 177–184.
680 A. SHARMILA AND P. MAHALAKSHMI
Downloaded by [Vellore Institute of Technology], [Sharmila ashok] at 21:08 15 November 2017