Available online at [Link].

cn

Open Access at PubMed Central

The Journal of Biomedical Research, 2020 34(3): 149–159

Original Article

EEG epileptic seizure detection and classification based on

dual-tree complex wavelet transform and machine learning
algorithms
Itaf Ben Slimen1,✉, Larbi Boubchir2, Zouhair Mbarki1, Hassene Seddik1
1Centre de Recherche et de Production Research Lab., Ecole Nationale Supérieure des Ingénieurs de Tunis, University
of Tunis, Tunis 1008, Tunisia;
2Laboratoire d'Informatique Avancée de Saint-Denis Research Lab., University of Paris 8, Saint-Denis, Cedex 93526,

France.

Abstract
The visual analysis of common neurological disorders such as epileptic seizures in electroencephalography
(EEG) is an oversensitive operation and prone to errors, which has motivated the researchers to develop effective
automated seizure detection methods. This paper proposes a robust automatic seizure detection method that can
establish a veritable diagnosis of these diseases. The proposed method consists of three steps: (i) remove artifact
from EEG data using Savitzky-Golay filter and multi-scale principal component analysis (MSPCA), (ii) extract
features from EEG signals using signal decomposition representations based on empirical mode decomposition
(EMD), discrete wavelet transform (DWT), and dual-tree complex wavelet transform (DTCWT) allowing to
overcome the non-linearity and non-stationary of EEG signals, and (iii) allocate the feature vector to the relevant
class (i.e., seizure class "ictal" or free seizure class "interictal") using machine learning techniques such as support
vector machine (SVM), k-nearest neighbor (k-NN), and linear discriminant analysis (LDA). The experimental
results were based on two EEG datasets generated from the CHB-MIT database with and without overlapping
process. The results obtained have shown the effectiveness of the proposed method that allows achieving a higher
classification accuracy rate up to 100% and also outperforms similar state-of-the-art methods.

Keywords: electroencephalography, epileptic seizure detection, feature extraction, dual-tree complex wavelet
transform, machine learning

Introduction epilepsy diagnosis, such as electroencephalogram

(EEG) signals recorded from the patient's scalp, which
According to the World Health Organization, is a collection of potential signals from many
epilepsy affects about 1% to 2% of the world's electrodes and they represent the spatial distribution of
population; about 65 million people have epilepsy[1–2]. the potential fields in the brain. EEG is an invaluable
Different examinations are used to confirm the tool for characterization of the spatial-temporal

✉Corresponding author: Itaf Ben Slimen, Electric Engineering de- CLC number: R742.1, Document code: A
partment, Centre de Recherche et de Production Research Lab, The authors reported no conflict of interests.
Ecole Nationale Supérieure des Ingénieurs de Tunis, University of
Tunis, street Taha Hussein Montfleury, Tunis 1008, Tunisia. Tel: This is an open access article under the Creative Commons Attribu-
+216-27-542-572, E-mail: itafslimen@[Link]. tion (CC BY 4.0) license, which permits others to distribute, remix,
adapt and build upon this work, for commercial use, provided the
Received 15 February 2019, Revised 30 November 2019, Accepted original work is properly cited.
11 February 2020, Epub 24 April 2020

© 2020 by the Journal of Biomedical Research. [Link]

150 Ben Slimen I et al. J Biomed Res, 2020, 34(3)

dynamics of neuronal activity in the brain; hence, used techniques for extracting EEG signal
diagnosis by manual observation of these signals is a characteristics. For the same purpose, the DTCWT
very sensitive task that resulted in practical tools for technique[20], applies filter banks to handle the wavelet
the research area dedicated to the diagnosis of seizure domain. This technique was recently applied for
detection[3]. epileptic seizure detection which demonstrated its
More than 120 million articles cited for epileptic performance for extracting the characteristics of the
seizure detection proved that this subject is still topical time-frequency domain of EEG signals[21]. After
and should be addressed. Various epileptic seizure extracting features, the next step is to employ
detection methods have been proposed in literature, classification methods in order to determine the class
and a brief review of some techniques related to our of input features. Various classification methods are
work was conducted and introduced in the next available. These range from rule-based decision-
section. Gotman and Gloor were among the first to making[4] and SVM[22] to artificial neural network
detect seizures using EEG signals[4]. The first machine (ANN). For example, the classification of EEG signals
learning approach was developed using the CHB-MIT was investigated using k-means classifier and
database[5]. Some researchers have used a large multilayer perceptron neural network model[14]. A
database, extracted from Freiburg and CHB-MIT supervised automatic learning method was proposed
databases for seizure detection, and their proposed for detecting seizures using linear discriminant
method allows achieving a high detection performance analysis (LDA)[23].
up to 100%[6]. Some proposed methods are based on The goal of this work is to build a reliable and
artificial neural networks allowing achieving a perfect accurate seizure detection method to facilitate and
classification rate for detecting epileptic seizures[7]. accomplish the diagnosis of epilepsy. Artifact
Some are based on support vector machine (SVM) eliminator is a valuable aid for diagnosis in EEG
with time-frequency features and wavelet variance signal, using the Savitzky-Golay filter and the
features allowing achieving correct classification rates MSPCA. To overcome the non-linearity and non-
up to 100%[8]. Another seizure detection method is stationary of EEG signals, the time-frequency
based on dual-tree complex wavelet transformation techniques have been developed to eliminate the strict
(DTCWT) and Fourier features allowing achieving a restrictions of classical frequency methods. Three
correct classification rate up to 100%[9]. popular signal decomposition techniques chosen in
Various approaches for epileptic seizure detection this work include empirical mode decomposition
have been proposed in literature. The processing of (EMD), DWT, and DTCWT. After crossing through a
EEG signals requires that these data must be noiseless step of denoising the artifacts and extracting the
to establish reliable disease diagnosis; therefore, a frequency bands using EMD, DWT, and DTCWT,
filter step is necessary for this purpose. Some statistical values were used to extract the relevant
researchers proposed to use the Savitzky-Golay filter characteristics of these frequency bands. Then three
and have its performance verified[10], while others classifiers k-nearest neighbor (k-NN, LDA, and SVM
proposed a multi-scale principal component analysis were established to distinguish between classes (i.e.,
(MSPCA)[11]. In general, all had adopted the interictal and ictal).
objectives of the extraction characteristics and their The choice of MSPCA and Savitzky-Golay for
classification, in order to differentiate between classes artifact removal is justified by their proven superiority
(i.e., seizure free: "interictal" class and seizure: "ictal" in different biomedical signals kinds, such as ECG[24],
class). Frequency domain analysis[11] and other time- EMG[25], and EEG[26]. The plethora of EMD, DWT,
frequency approaches, such as the short-term Fourier and DTWCT signal decomposition methods used in
transform[12] and the wavelets[13,14–15] have often been different fields of analysis justifies the choice in this
used to extract discriminant characteristics work. This step is followed by the extraction of the
representing the EEG signals. Advanced detection of statistical characteristics to extract the necessary
epileptic seizures used a feature extraction technique information with reduction of used data dimensions.
based on wavelets[16]. Some features have shown high The aforementioned modulators were combined with
performance for characterization changes in EEG three popular machine-learning techniques to find a
signals such as entropy, energy[17–18], principal powerful model for detecting epileptic seizures.
component analysis (PCA)[19], and independent Therefore, this work aims to develop an EEG signal
component analysis (ICA). classification system for seizure detection using CHB-
Based on these previous studies, the discrete MIT (EEG scalp) databases, and to ensure the
wavelet transform (DWT) is one of the most widely robustness of the proposed model in which two
 EEG-based epilepsy detection using machine learning 151

databases have been extracted and realized based on is needed to clean the data. Many signal-filtering
EEG segments with and without overlapping algorithms have been proposed to improve the
procedure. The results proved that the proposed model performance of health devices (EEG, ECG, EMG...)
is suitable for epileptic seizure detection. Fig. 1 besieged by artifacts. Savitzky-Golay filter and multi-
illustrates the work plan presented in this study. scale principal component analysis have been used in
recent research and their performance has been
Materials and methods verified.
Savitzky-Golay filter
The lack of visual difference between the free
seizure and seizure EEG signals makes the epileptic During the pre-processing of biomedical signals,
seizure detection a difficult task. Therefore, three the most challenging problem is to extract high
proposed steps were employed, with EEG signal resolution EEG signals from noisy measurements and
denoising techniques followed by features extraction retain the EEG waveform sharpness. The spectral
methods to enable faster and more effective seizure quality of a signal [signal-to-noise (S/N) ratio] may be
improved by increasing filter width or smoothing the
detection. Hence, the classification using the extracted
signal multiple times. Hence, The Savitzky-Golay
characteristics is performed using different machine
filter is a low-pass filter used for attenuation noise,
learning. The EEG database used in our study is
without damaging or destroying the spectral properties
presented with a brief description.
of data. This filter is defined as a weighted moving
EEG database average that performs a least-squares-fit convolution
procedure.
All tests and simulations were conducted based on Therefore, the S –Golay algorithm is characterized
the database collected at the Boston Children's by the selected data interval, in other words, the
Hospital (CHB) and used to validate the experimental window size, a polynomial function is fitted to the
results. This "CHB-MIT" database includes 23 chosen size window and the data point (noisy signal)
different subsets containing EEG records of pediatric is calculated using the polynomial coefficients. The
patients. It contains 182 seizures. All signals were set of data n{x j , y j } points, where j=1,2, …,n; x is an
sampled at 256 Hz per second with a resolution of 16 independent variable and y is the observed value, can
bits. Most files contain 23 EEG signals. The be represented with convolution coefficients called c
international 10-20 EEG electrodeposition and and Ci and expressed by:
nomenclature system was used for these recordings[5].
i= c−1
∑ 2
c+1 c−1
EEG signal pre-processing Yj = Ci y j+1 ⩽ j ⩽ n− (1)
2 2
i=− c−1
Biomedical signals such as EEG were contaminated 2

by artifacts and noise, and thus a pre-processing step Hence, S-Golay filter requires the inputs which
Sub-bands
EEG signal Decomposition decomposition
DWT

EMD
Pre-processing
DTCWT
S-Golay

MSPCA
Feature extraction

Classification
No seizure segment SVM

k-NN
Seizure segment
LDA

Fig. 1 Diagram for proposed seizure detection approach. MSPCA: multi-scale principal component analysis; DWT: discrete wavelet
transform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet transformation; SVM: support vector machine; k-NN:
k-nearest neighbor; LDA: linear discriminant analysis.
152 Ben Slimen I et al. J Biomed Res, 2020, 34(3)

include: the input noisy signal (x), the polynomial approaches. These, such as DWT, DTCWT, and
order (k) and its frame size (f). One of the advantages EMD, are recently used in the literature and will be
of this filter is to retain distribution characteristics, used in this work and represented in this section with
which are often flattened with other smoothing their proven superiority to reach this goal.
methods[10,27].
Empirical mode decomposition (EMD)
Multi-scale principal component analysis (MSPCA)
EMD is a procedure used to decompose signals into
The MSPCA approach is proposed to achieve the different simple oscillatory modes called intrinsic
objective of eliminating noise without sacrificing a function mode (IMF). Hence, the signal x(t) can be
significant amount of fast changes in the signal and to represented by the sum of these IMFs components and
provide a dimension reduction with select components by a residue as follows[29]:
relevant in the different scales. The advantages of this
∑
M
method come from the combination of the PCA, x (t) = Dm (t) + r M (t) (2)
which is responsible for extracting the relationship m=1
between several variables, and the wavelets, that where M is the number of IMFs and rM(t) is the
divide the stochastic processes of the determinists and residue. The IMFs of a signal must satisfy two
approximately decorrelate the auto-correlation bet- necessary conditions. Firstly, the difference between
ween the variables. The MSPCA calculates the PCA the total numbers of extrema and zero-crossings is at
of all wavelet coefficients for all wavelet decomposition most one. Secondly, in all points, the average value of
levels, and then establishes a relationship between the the envelope defined by the local maxima and that by
results obtained and the suitable levels[24]. Fig. 2 the local minima is zero or very close to zero[30]. The
illustrates the steps of the MSPCA method and the following steps represent the EMD algorithm:
following algorithm explains it: 1) Extract local maxima and local minima from
1) Perform wavelet decomposition at level I for signal of x(t),
each column of data matrix X. 2) Connect between all minima and all maxima to
2) For 1≤i≤I, perform the PCA of the detailed obtain two envelopes emin(t) and emax(t),
matrix GiX, choose an appropriate number of 3) Compute the average between them as:
significant principal components pi or suppress the [emin (t) + emax (t)]
detail. m (t) = ,
2
3) Perform the PCA of the approximation matrix 4) Extract the detail from x(t) as: d(t)=x(t)-m(t)
HIX and select a suitable number of principal 5) Examine, if d(t) is an IMF or not by verifying the
components pI+1. above conditions for IMFs,
4) The main characteristics of the original matrix X 6) Reiterate on the residual d(t).
put in a new matrix Ẋ , by inverting the DWTT wavelet After having the first IMF D1(t) of the signal, to get
transformation. other IMFs, generate the residual r(t) as follows:
5) Finally, execute the PCA of Ẋ [28]. r(t)=x(t) – D1(t). The process described above will be
continued until IMF can no longer be obtained from
EEG signal decomposition
the final residue.
To characterize the EEG signal and determine its
Discrete wavelet transform (DWT)
class, some methods have often been used to extract
discriminant characteristics representing EEG signals DWT is one of the most efficient time-frequency
from frequency analysis and time-frequency tools for non-stationary signal analysis. The DWT
THRESHOLDING

PCA (H1X)

PCA (G1X) ·
X DWT DWTT X
PCA
PCA (G1X)

PCA (G1X)

Fig. 2 The procedure for MSPCA. DWT: discrete wavelet transform; PCA: principal component analysis.
 EEG-based epilepsy detection using machine learning 153

procedure makes it possible to present an input signal h0 2

h0 2
in sets of functions, called wavelets, by scaling and
h0(1) 2 h1 2
shifting the mother function into wavelets, that is to
h1 2
say, to decompose a signal into two components g0 2
h1(1) 2
called sub-bands: a low frequency component called s g0 2
"approximation coefficients A", and a high frequency g0(1) 2 g1 2
component called "detailed coefficients D". The low g1 2
frequency component can further be decomposed into g1(1) 2
approximations and details. After each decomposition,
the frequency resolution is doubled and the temporal Fig. 4 An illustrative scheme of the DTCWT at level 3.
resolution is halved. After choosing the decomposition
level j, the components of the signal s(k) can be Consequently, the corresponding wavelets ψh(t) and
represented by the following expression and Fig. 3[31]: ψg(t) form approximately a Hilbert pair, while the
∑ scaling function φg(t) is approximately the Hilbert
D j (i) = s (k) · h(2 · i − k) (3) transform of φh(t).
k
∑ Discriminative feature extraction
A j (i) = s (k) · l(2 · i − k) (4)
k The classification requires a necessary step whose
Regarding Fig. 3, h and g are low-pass and high- characteristic extraction aims to obtain the most
pass filters, respectively. s is the input original signal important information and preserve the small size of
where Ca1, Ca2, and Ca3 represent the approximation the data. Therefore, after signal decomposition using
the DWT, EMD and DTCWT methods, the most
coefficients of h(1), h(2) and h(3) and Cd1, Cd2 and
prominent statistical features have been used for this
Cd3 represent the detail coefficients of g(1), g(2) and
study to obtain the class of the input EEG signal:
g(3) at level 1, 2 and 3, respectively.
Mean of coefficients in each sub-band:
Dual-tree complex wavelet transform (DTCWT)
1 ∑
M

DTCWT, a DWT enhancement, uses double µ= yj (7)

M j=1
wavelet filters low and high at each scale to obtain
two parts: a real and an imaginary complex wavelet Average power of coefficients in each sub-band:
coefficient. This transformation is invariant by v
u
t
1 ∑ 2
M
directional and selective shifting in two or more λ= y (8)
dimensions, which are very important in applications M j=1 j
such as pattern recognition and signal analysis.
Standard deviation of coefficients in each sub-band:
Hence, the DTCWT can be represented by the
v
u
t
1 ∑(
following expression[20]: M )2
σ= yj −µ (9)
ψ (t) = ψh (t) + iψg (t) (5) M j=1
φ (t) = φh (t) + iφg (t) (6)
Signal classification
where the complex wavelet is ψ(t) while the
complex scaling function is φ(t). With reference to Frequency band decomposition is performed to
Fig. 4, s represents the input signal, and the output represent an EEG signal in feature vector and to
coefficients of the top tree h and those of the bottom obtain the most important information of the data to
tree g are considered as the real and imaginary parts, determine its class. The identification of signals relied
respectively. Therefore, h(1) (1)
0 , h0 , g0 , and g0 are a low-
on classification algorithms to estimate the desired
(1) (1) class. For this purpose, three popular classifiers are
pass filter and h1 , g1 , h1, and g1 are high-pass filters.
used, namely LDA, k-NN, and SVM, which are
h (3) 2 Ca3 briefly explained in this section.
h (2) Ca2
2
Ca1 Cd3 Linear discriminant analysis (LDA)
h (1) 2 g (3) 2
s g (2) 2 Cd2 LDA, also called Fisher's LDA, is a simple
Cd1
g (1) 2 technique used in machine learning, statistics, and
pattern recognition to find a linear separating two or
Fig. 3 DWT decomposition tree. more classes, and in general, it gives a good class
154 Ben Slimen I et al. J Biomed Res, 2020, 34(3)

estimation. Fisher's LDA requires very little second based on overlapping process. The segment
computation. LDA aims to use hyperplanes to separate extraction process of these two datasets is clearly
data from different classes, and the separator explained in this section. The performance evaluation
hyperplane is obtained by looking for the projection criteria to assess the performance and the
that maximizes the distance between classes and effectiveness of the proposed method and the
minimizes the interclass variance between them. For a experimental results obtained will be shown as
class number N (N>2), several hyperplanes are follows.
used[32–33]. In this work, the linear discriminant type The developed seizure detection model is based on
was used for regularized linear discriminant analysis. EEG segment extraction with an 8-second rectangular
window. The seizure time is known and given in the
Support vector machine (SVM)
description database for each patient. The first
With excellent performance results, SVM is applied extracted dataset contains 400 EEG segments
in various fields. Contrary to LDA, which constructs a extracted approximately from all the recorded signals
probabilistic model for each class using all data in the CHB-MIT database with 200 no seizure
points, SVM is to minimize a linear separation error segments ("interictal") and 200 seizure segments
by focusing on the neighboring points of the linear ("ictal").
separation surface, that is to say, the points located far According to a previous study[6], the choice of the 8
from the separation surface are ignored in the learning seconds window size (with 2 048 samples) provides a
process. SVM is based on the Vapnik-Chervonenkis more robust performance than that of shorter or longer
theory and the principle of structural risk windows. To prove the effectiveness of the proposed
minimization. Its purpose is to determine the seizure detection model, the overlapped database has
minimizing of the training set error by maximizing the been created with an overlapping process based on the
boundary of the separating hyper-plane between the first extracted database that will be presented as
data[34–35]. In this work, the Kernel function used to follows.
differentiate between the two classes is linear kernel
Overlapping procedure
with the Kernel scale parameter chosen to be 1.
While the first database was created (with 400
k-nearest neighbor (k-NN)
segments), to evaluate the performance and better
The k-NN is a machine-learning technique used for prove the effectiveness of the proposed seizure
the classification task, which aims to find the nearest detection method, another database will be created
k-neighbors among the learning set, and the categories based on the first realized with an overlapping
of k closest neighbors, which are used to weight the process. Hence, new segments will be created by
candidates in the category. The closest neighbors to overlapping between two segments belonging to the
the sample are expressed as distance such as same EEG signal in the same class. Let suppose that:
Euclidean distance, Chebyshev distance, and - N is the segment number of the first database
Manhattan distance. The performance of k-NN already created (without overlapping) called B. N1 and
algorithm strongly depends on two factors: an N2 is the number of segments in "interictal" class B1
appropriate similarity function and an appropriate and "ictal" class B2 respectively,
value for k. If k is too large, large classes will - Nc is the segment number of the second database
overwhelm small classes, whereas if k is too small, the with overlapping process called Bc. Nc1 and Nc2 are
advantage of the k-NN algorithm cannot be the numbers of segments in "interictal" class Bc1 and
traced[36–37,19]. In this research, 3-NN algorithm (3 "ictal" class Bc2 respectively.
nearest neighbors) and Euclidean distance metric were Therefore, the overlap process is presented as
used to classify the EEG signals. follows:
- n represents the segment samples number (2 048),
Database preparation
i is segment index (ith segment), p is the point offset
The CHB-MIT database is one of the EEG for overlap, fe is the sampling frequency 256 Hz, and
databases used in recent research for epileptic seizure sc is the segment seconds number (8 seconds).
detection. To prove the effectiveness of the proposed Example: let suppose i and i+1 are two segments of
method, two EEG datasets have been extracted from the same class to be overlapped where X1i , X2i ,
the CHB-MIT database. Hence, the extraction of X3i , X4i , X5i , X6i , X7i , X8i and X1i+1 , X2i+1 , X3i+1 ,
shorter EEG segments based on no overlapping X4i+1 , X5i+1 , X6i+1 , X7i+1 , X8i+1, are represent the
process represents the first used database, while the sample of segments of i and i + 1, respectively, and j
 EEG-based epilepsy detection using machine learning 155

the new obtained segment with a gap between the extracted features was considerably doubled in
segment is p = 6 . DTCWT, the performance in the testing step
i = [X1i , X2i , X3i , X4i , X5i , X6i , X7i , X8i ] improved.

i+1 = [X1i+1 , X2i+1 , X3i+1 , X4i+1 , X5i+1 , X6i+1 , X7i+1 , X8i+1 ] Performance evaluation criteria

j = [X3i , X4i , X5i , X6i , X7i , X8i , X1i+1 , X2i+1 ] To prove the robustness of the proposed method,
cross-validation (CV) is a procedure mainly applied to
Finally, the new segment can be presented by the
estimate the proficiency of machine learning
following expression:
{[n ] [ ]} algorithm. CV is used to estimate how the model
n should work when used to make predictions of unused
j= i × (sc − p) : n × sc ; i + 1 1 : × (sc − p)
sc sc data at the step of preparation and model learning. In a
(10)
10 times CV, the database is randomly separated into
{[ ] [ ]}
j = i f e × (sc − p) : n × sc ; i + 1 1 : f e × (sc − p) 10 different mutually exclusive folds of the same size,
(11) where nine folds will be used to training the data and
The size of new database (overlapped) will be the remaining one will be used to test and predict the
increased compared to the first database. classification. The CV accuracy is expressed by:
N = N1 + N2 and Nc = Nc1 + Nc2 ; hence, Nc1 =
1∑
k
(N1 × 2) − 1 and Nc2 =(N2 × 2) − 1. CV = Ai (13)
k i=1
Nc = [(N1 × 2) − 1] + ((N2 × 2) − 1) = 2N1 + 2N2 − 2
(12) where represents the number of folds used, and Ai is
Therefore, the overlapped database size will be 798 the accuracy measure of each, i = 1, . . . , k. In addition,
segments, with 399 segments in both classes and each there are three performance measures to evaluate
length segment 8 seconds. The experimental results of machine learning methods using true positives (TP),
the proposed approach using the two extracted false positives (FP), true negatives (TN), and false
databases will be presented by following the negatives (FN), which can be expressed by:
performance criteria to validate the model accuracy. TP
Sensitivity (%) = × 100 (14)
T P + FN
Feature selection process
TN
The sampling rates of the databases are 256 Hz; the Speci f icity (%) = × 100 (15)
T N + FP
polynomial order was chosen to be 2 on S-Golay filter
and the frame length was selected to be 17, proved to TP+TN
Accuracy (%) = × 100 (16)
be the parameter' choice[10]. In MSPCA, a five-level T P + FN + T N + FP
decomposition was used to generate the sub-bands
signals. The Kaiser rule was responsible for Results
eliminating the number of loadings[26]. After the
artifact removal procedure with S-Golay and MSPCA, The previous sections have shown that the first goal
the features have been extracted from the sub-band in this work is to find a model to classify EEG signal
decomposed signals using EMD, DWT, and DTCWT segments to develop automatic seizure detection
methods. The Daubechies wavelets of different orders systems using the children's EEG database CHB-MIT.
(2, 3, 4, 5, and 6) are investigated for the analysis of However, shorter EEG segments with 8 seconds have
epileptic EEGs. This family of wavelets has been extracted from this database that makes it
orthogonal property and is efficient as a filter possible to develop a practical and robust seizure
implementation[38]. The fourth order Daubechies detection system to diagnose epilepsy. Ictal record
wavelet is found to be most appropriate for analysis of signals represent a small amount of data compared to
EEG data[6,39]. Hence, DB4 wavelet function was used interictal record signals. Therefore, a subsampling
to decompose the EEG signals with DWT. The paradigm of the class with larger proportions
decomposition levels in all methods were selected to (interictal) is used to circumvent the imbalance
be 3. Hence, 4, 8 and 3 sub-band signals were between the two classes[40], in order to have better
obtained using DWT, DTCWT, and EMD accuracy of seizure detection. In this work, segments
respectively. Therefore, in each sub-band signals three of interictal data were randomly selected to produce a
statistical characteristics have been extracted to obtain comparable and balanced number of ictal and
12, 24, and 9 features for DWT, DTCWT, and interictal segments.
respectively. However, although the number of Based on the segment model, the purpose of this
156 Ben Slimen I et al. J Biomed Res, 2020, 34(3)

study was to develop an EEG signal classification filter in the removal of artifacts; MSPCA achieved a
model. Therefore, the contribution is to have a 100% accuracy with the three classification methods
powerful and suitable model using three popular while S-Golay 98%.
methods to extract discriminant characteristics such as With the three EEG signal decomposition methods
EMD, DWT, and DTCWT, followed by a dimension compared, the results indicate that using the two
reduction employing three different statistical denoising methods, DTCWT was superior to DWT
characteristics and three commonly used methods of and much better than EMD where with S-Golay
automatic learning such as LDA, k-NN, and SVM. (DTCWT, 99%; DWT, 94%; and EMD, 91%). Fig. 5
To prove the performance results of the average also shows that the DTCWT gave a fairly high
precision, extraction methods were evaluated and the
classification result with 100% accuracy classification
results are presented in detail and illustrated in Fig. 5.
rates, which is attributed to the approximate shift
The parameter values of all selected techniques are
invariant property, a very important factor in epileptic
explained in the preview section because they produce
seizure detection. On aspects of the performance of
the most robust results. In Fig. 5, the plots are shown
in the left column present the S-Golay results to apply feature extraction methods, the experimental results
a series of data points to denoise the signals without show the superiority of the standard deviation and the
deforming its input, while the plots are shown in the average power, which both gave a rather high
right column present the MSPCA results, which classification accuracy when compared to the mean of
shows the advantage of PCA in extracting necessary the sub-band features. Hence the mean static feature
information and the effectiveness of extracting with DTCWT outperforms that with DWT, which
deterministic features from stochastic processes using ensures the high effectiveness of DTCWT. In
wavelets. With the two filter processes compared, it addition, the result provides that k-NN and SVM gave
can be found that MSPCA outperformed the S-Golay 100% classification accuracy when compared to LDA.

DTCWT+S-GOLAY DTCWT+MSPCA
k-NN LDA SVM k-NN LDA SVM
Classification accuracy (%)

Classification accuracy (%)

99 93 99 97 99 98 99 100 99 100 91 100 96 98 100 96 100

100 100
77
80 80
60 60
40 40
20 20
0 0
Average power Mean Standard deviation Average power Mean Standard deviation

DWT+S-GOLAY DWT+MSPCA
k-NN LDA SVM k-NN LDA SVM
Classification accuracy (%)

Classification accuracy (%)

94 88 92 98 98 98 93 88 92 100 87 100 94 100 96 100

100 100 91
75
80 80
60 60
40 40
20 20
0 0
Average power Mean Standard deviation Average power Mean Standard deviation

EMD+S-GOLAY EMD+MSPCA
k-NN LDA SVM k-NN LDA SVM
Classification accuracy (%)
Classification accuracy (%)

93 96 100 100 100 94 100

100 91 82 88 90 100 85 90 92
79 81 75
80 65 80
60 60
40 40
20 20
0 0
Average power Mean Standard deviation Average power Mean Standard deviation

Fig. 5 Classification accuracy of the proposed method using the signal decomposition transforms for feature extraction considered
in this study (i.e., EMD, DWT, DTCWT) with EEG signal pre-processing techniques: MSPCA (right column) and S-Golay (left
column).
 EEG-based epilepsy detection using machine learning 157

Discussion invariance. The DTCWT has solved these problems

by its non-oscillating property with approximate
The overall results obtained in this study show that invariant shift amplitude where this property is very
combination of the DTCWT with SVM and k-NN important for the detection of epileptic seizures in
improves the accuracy rates of the classification EEG signals.
especially by using the standard static deviation and In conclusion, we have proposed a robust automatic
average power characteristics with the two denoising approach for epileptic seizure detection and
processes. As shown in Fig. 5 which (MSPCA+ classification that was assessed on two EEG datasets
DTCWT+STD+SVM), (MSPCA+DTCWT+STD+k- generated from the CHB-MIT database. The proposed
NN, (S-Golay+DTCWT+STD+SVM) and (S-Golay+ approach is based on two denoising signal methods
DTCWT+STD+k-NN achieve a 100% classification (i.e., S-Golay and MSPCA) combined with three
accuracy rate. Hence, it can be concluded that the popular feature extraction techniques (i.e., DTCWT,
proposed model is reliable and robust for seizure DWT, and EMD). Three different machine-learning
detection with the DTCWT combined with SVM and techniques (namely SVM, k-NN, and LDA) have been
k-NN. used to distinguish between the seizure-free and
To compare the results of the proposed approach seizure segments based on the feature extraction
with seizure detection methods realized in literature, methods already mentioned. The experimental results
Table 1 provides that the proposed approach has have shown that the employment of DTCWT has been
achieved an overall perfect accuracy of 100% in very successful because it outperforms the other two
methods (DWT and EMD). The results obtained have
seizure detection using EEG segments.
also clearly shown that MSPCA+DTCWT+SVM and
The obtained results using the overlapped database
k-NN offers a perfect classification with 100%
ensure the reliability and stability of this seizure
accuracy. The findings demonstrate the effectiveness
detection model as illustrated in Table 2 and Fig. 6.
of the proposed method compared to existing
The combination of DTCWT and MSPCA offers
methods. After completing the seizure detection task,
advantages over the DWT and EMD and keeps the
future studies may take more interest in the prediction
same 100% accuracy performance. The important
of epileptic seizures.
difference between EMD and wavelet-based methods
lies in the process of feature extraction. IMF signals
are the product of EMD methods and have the same
References
length as the original EEG signals, which have given
no reduction in the number of samples. On the other
[1] Acharya UR, Sree SV, Swapna G, et al. Automated EEG
hand, the coefficients characterizing different sub-
analysis of epilepsy: a review[J]. Knowl-Based Syst, 2013, 45:
bands produced by wavelet-based methods (DWT and 147–165.
DTCWT) have a much smaller number of samples [2] Moshé SL, Perucca T, Ryvlin P, et al. Epilepsy: new
than the EEG signals input. However, the DWT advances[J]. Lancet, 2015, 385(9971): 884–898.
suffers from certain disadvantages such as the [3] Adeli H, Ghosh-Dastidar S. Automated EEG - based diagnosis
oscillatory nature, the aliasing of the spectrum with of neurological disorders inventing the future of neurology[M].
limited directional information, and the lack of lag New York: CRC Press, 2010: 71–75.

Table 1 Comparison of the proposed method against some state-of-the-art methods

Method Feature extraction Classifier Accuracy

Fergus et al[23] (2014) Band-pass filter, peak and median frequency, RMS, Entropy based-features k-NN 88%

Raffiudin et al [16] (2011) Wavelet-based features+statistics-based features+IQR+MAD LDA 96%

Alickovic et al (2018)[6] Multiscale PCA, EMD, DWT, WPD, statistics-based features RF, SVM, MLP, and k-NN 99.7%

Swami et al (2016)
[21] DTCWT GRNN 98%

Gandhi et al [18] (2011) DWT, energy, standard deviation and entropy based-features SVM and PNN 99%

Approximate entropy, sample entropy, HOS, decision tree, fuzzy-based GMM, k-NN, SVM, and
Acharya et al[13] (2012) 99%
features RBPNN

Swami et al[22](2014) WPT, energy, standard deviation and entropy based-features SVM 99%

Proposed method MSPCA, DTCWT SVM and k-NN 100%

MSPCA: multi-scale principal component analysis; DWT: discrete wavelet transform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet
transformation; SVM: support vector machine; k-NN: k-nearest neighbor; LDA: linear discriminant analysis.
158 Ben Slimen I et al. J Biomed Res, 2020, 34(3)

Table 2 Classification accuracy of the proposed overlapping method

Features

Overlapping Average power Mean Standard deviation

2 4 6 2 4 6 2 4 6

S-Golay 98 98 96 75 80 78 97 98 96
SVM
MSPCA 100 100 100 87 71 71 100 100 100

S-Golay 98 97 96 86 88 89 99 98 96
DTCWT k-NN
MSPCA 100 100 100 93 92 93 100 100 100

S-Golay 93 92 93 70 81 71 99 93 92
LDA
MSPCA 91 87 91 76 74 75 98 96 97

S-Golay 94 92 92 75 78 75 93 94 94
SVM
MSPCA 100 100 99 76 71 72 100 100 99

S-Golay 95 93 96 84 89 90 95 92 96
DWT k-NN
MSPCA 100 100 99 88 90 91 100 100 100

S-Golay 98 90 89 80 81 81 91 91 92
LDA
MSPCA 90 91 90 76 72 75 98 93 98
S-Golay 94 95 95 66 61 64 93 94 92
SVM
MSPCA 100 100 99 65 70 63 100 100 99

S-Golay 94 94 94 75 76 84 96 92 94
EMD k-NN
MSPCA 99 100 99 92 91 78 99 100 100

S-Golay 80 81 83 50 62 60 88 90 91
LDA
MSPCA 84 88 87 63 72 67 99 97 97
MSPCA: multi-scale principal component analysis; DWT: discrete wavelet transform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet
transformation; SVM: support vector machine; k-NN: k-nearest neighbor; LDA: linear discriminant analysis.

100
Sensitivity or specificity (%)

90
80
70
60
50
40 Average power
30 Mean
20 Standard deviation
10
0
MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA

MSPCA
S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

S-Golay

Sens Spec Sens Spec Sens Spec Sens Spec Sens Spec Sens Spec Sens Spec Sens Spec Sens Spec
SVM LDA k-NN SVM LDA k-NN SVM LDA k-NN
DWT EMD DTCWT

Fig. 6 The sensitivity and specificity of the proposed method using the signal decomposition transforms for feature extraction con-
sidered in this study (i.e., EMD, DWT, DTCWT) with EEG signal pre-processing techniques (MSPCA and S-Golay). MSPCA: multi-
scale principal component analysis; DWT: discrete wavelet transform; EMD: empirical mode decomposition; DTCWT: dual-tree complex
wavelet transformation; SVM: support vector machine; k-NN: k-nearest neighbor; LDA: linear discriminant analysis; Sens: sensitivity; Spec:
specificity.

[4] Gotman J. Automatic detection of seizures and spikes[J]. J Clin empirical mode decomposition, discrete wavelet transform, and
Neurophysiol, 1999, 16(2): 130–140. wavelet packed decomposition for automated epileptic seizure
[5] Shoeb AH. Application of machine learning to epileptic seizure detection and prediction[J]. Biomed Signal Process Control,
onset detection and treatment[D]. Cambridge: Harvard-MIT 2018, 39: 94–102.
Division of Health Sciences and Technology, 2009: 157–162. [7] Tzallas AT, Tsipouras MG, Fotiadis DI. Automatic seizure
[6] Alickovic E, Kevric J, Subasi A. Performance evaluation of detection based on time-frequency analysis and artificial neural
 EEG-based epilepsy detection using machine learning 159

networks[J]. Comput Intell Neurosci, 2007, 2007: 80510. [23] Fergus P, Hignett D, Hussain AJ, et al. An advanced machine
[8] Xie SK, Krishnan S. Wavelet-based sparse functional linear learning approach to generalised epileptic seizure
model with applications to EEGs seizure detection and epilepsy detection[C]//Proceedings of the 10th International Conference
diagnosis[J]. Med Biol Eng Comput, 2013, 51(1-2): 49–60. on Intelligent Computing. Taiyuan, China: 2014, Springer:
[9] Chen GY. Automatic EEG seizure detection using dual-tree 112–118.
complex wavelet-Fourier features[J]. Expert Syst Appl, 2014, [24] Alickovic E, Subasi A. Effect of Multiscale PCA de-noising in
41(5): 2391–2394. ECG beat classification for diagnosis of cardiovascular
[10] Acharya D, Rani A, Agarwal S, et al. Application of adaptive diseases[J]. Circuits, Syst Signal Process, 2015, 34(2):
savitzky-golay filter for EEG signal processing[J]. Perspect 513–533.
Sci, 2016, 8: 677–679. [25] Gokgoz E, Subasi A. Effect of multiscale PCA de-noising on
[11] Polat K, Günes S. Classification of epileptiform EEG using a EMG signal classification for diagnosis of neuromuscular
hybrid system based on decision tree classifier and fast Fourier disorders[J]. J Med Syst, 2014, 38(4): 31.
transform[J]. Appl Mathem Comput, 2007, 187(2): 1017–1026. [26] Kevric J, Subasi A. The effect of multiscale PCA de-noising in
[12] Duque-Muñoz L, Espinosa-Oviedo JJ, Castellanos-Dominguez epileptic seizure detection[J]. J Med Syst, 2014, 38(10): 131.
CG. Identification and monitoring of brain activity based on [27] Agarwal S, Rani A, Singh V, et al. EEG Signal enhancement
stochastic relevance analysis of short - time EEG rhythms[J]. using cascaded S-Golay filter[J]. Biomed Signal Process
BioMed Eng OnLine, 2014, 13: 123. Control, 2017, 36: 194–204.
[13] Acharya UR, Sree V, Ang PCA, et al. Application of non- [28] Aminghafari M, Cheze N, Poggi JM. Multivariate denoising
linear and wavelet based features for the automated using wavelets and principal component analysis[J]. Computat
identification of epileptic EEG signals[J]. Int J Neural Syst, Statist Data Anal, 2006, 50(9): 2381–2398.
2012, 22(2): 1250002. [29] Pachori RB, Patidar S. Epileptic seizure classification in EEG
[14] Gandhi TK, Chakraborty P, Roy PG, et al. Discrete harmony signals using second-order difference plot of intrinsic mode
search based expert model for epileptic seizure detection in functions[J]. Comput Methods Programs Biomed, 2014,
electroencephalography[J]. Expert Syst Appl, 2012, 39(4): 113(2): 494–502.
4055–4062. [30] Rilling G, Flandrin P, Gonçalvès P. On empirical mode
[15] Swami P, Gandhi TK, Panigrahi BK, et al. A comparative decomposition and its algorithms[C]//Proceedings of IEEE-
account of modelling seizure detection system using wavelet EURASIP Workshop on Nonlinear Signal and Image
techniques[J]. Int J Syst Sci: Oper Logist, 2017, 4(1): 41–52. Processing. Grado: IEEE, 2003.
[16] Rafiuddin N, Khan YU, Farooq O. Feature extraction and [31] Alickovic E, Subasi A. Medical decision support system for
classification of EEG for automatic seizure diagnosis of heart arrhythmia using DWT and random forests
detection[C]//Proceedings of 2011 International Conference on classifier[J]. J Med Syst, 2016, 40(4): 108.
Multimedia Signal Processing and Communication [32] Duda RO, Hart PE, Stork DG. Pattern classification[M]. 2nd
Technologies. Aligarh: IEEE, 2011. ed. New York: Wiley, 2001.
[17] Gandhi T, Panigrahi BK, Bhatia M, et al. Expert model for [33] Fukunaga K. Introduction to statistical pattern recognition[M].
detection of epileptic activity in EEG signature[J]. Expert Syst 2nd ed. San Diego: Academic Press, Inc, 1990.
Appl, 2010, 37(4): 3513–3520. [34] Boser BE, Guyon IM, Vapnik VN. A training algorithm for
[18] Gandhi T, Panigrahi BK, Anand S. A comparative study of optimal margin classifiers[C]//Proceedings of the Fifth Annual
wavelet families for EEG signal classification[J]. Workshop on Computational Learning Theory. Pittsburgh:
Neurocomputing, 2011, 74(17): 3051–3057. ACM, 1992.
[19] Subasi A, Gursoy MI. EEG signal classification using PCA, [35] Vapnik VN. The nature of statistical learning theory[M]. New
ICA, LDA and support vector machines[J]. Expert Syst Appl, York: Springer-Verlag, 1995.
2010, 37(12): 8659–8666. [36] Mitchell TM. Machine learning[M]. New York: McGraw-Hill,
[20] Selesnick WI, Baraniuk RG, Kingsbury NC. The dual-tree 1997.
complex wavelet transform[J]. IEEE Signal Process Mag, [37] Aha DW, Kibler D, Albert MK. Instance-Based learning
2005, 22(6): 123–151. algorithms[J]. Mach Learn, 1991, 6(1): 37–66.
[21] Swami P, Gandhi TK, Panigrahi BK, et al. A novel robust [38] Smit DJA, Boersma M, Schnack HG, et al. The brain matures
diagnostic model to detect seizures in with stronger functional connectivity and decreased
electroencephalography[J]. Expert Syst Appl, 2016, 56: randomness of its network[J]. PLoS One, 2012, 7(5): e36896.
116–130. [39] Metin A. Time frequency and wavelets in biomedical signal
[22] Swami P, Godiyal AK, Santhosh J, et al. Robust expert system processing[M]. New York: Wiley-IEEE Press, 1998: 174–176,
design for automated detection of epileptic seizures using SVM 207–210.
classifier[C]//Proceedings of 2014 International Conference on [40] Qiao XY, Liu YF. Adaptive weighted learning for unbalanced
Parallel, Distributed and Grid Computing. Solan: IEEE, 2014: multicategory classification[J]. Biometrics, 2009, 65(1):
219–222. 159–168.