Informatics in Medicine Unlocked 26 (2021) 100721

Contents lists available at ScienceDirect

Informatics in Medicine Unlocked

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

A comparison study of polynomial-based PCA, KPCA, LDA and GDA feature

extraction methods for epileptic and eye states EEG signals detection using
kernel machines
Laurent Chanel Djoufack Nkengfack a, b, *, Daniel Tchiotsop b, Romain Atangana c,
Beaudelaire Saha Tchinda b, Valérie Louis-Door d, Didier Wolf d
a
Unité de Recherche de Matière Condensée-d’Électronique et de Traitement Du Signal (UR-MACETS), Faculty of Science, University of Dschang-Cameroon, P.O. Box 67,
Dschang, Cameroon
b
Unité de Recherche d’Automatique et d’Informatique Appliquée (UR-AIA), IUT-FV of Bandjoun, University of Dschang-Cameroon, P.O. Box 134, Bandjoun, Cameroon
c
Division of Continuing Education and Distance Learning, Higher Teacher Training College (HTTC) of Bertoua, University of Ngaoundéré-Cameroon, P.O. Box 652,
Bertoua, Cameroon
d
Centre de Recherche en Automatique de Nancy (CRAN), UMR CNRS 7039, ENSEM de Lorraine, Nancy, France

A R T I C L E I N F O A B S T R A C T

Keywords: Background and objective: Initially, analysis of Electroencephalogram (EEG) signals was purely visual, tedious,
Polynomial transforms time-consuming, and required a physician. Changing this old approach to classification proves to be an
Feature extraction methods extraordinary task that gained much attention and a great deal of effort. With this intention, this comparison
Low-dimensional features
study focused on the development of polynomial-based feature extraction methods for epileptic and eye states
Kernel machines
Epileptic and eye states electroencephalogram
EEG signals detection using kernel machines.
(EEG) signals detection Method: Polynomial transforms are applied to decompose EEG signals in the frequency domain before their
analysis using linear and non-linear measures. Thereafter, the standard and kernel extension methods are applied
to determine principal components and discriminants which help to extract informative and discriminative low-
dimensional features. For direct detection of EEG signals, extracted features are fed into kernel machines namely
simple multilayer perceptron neural network (sMLPNN) and least-square support vector machine (LS-SVM).
Results: Using the publicly available Bonn-University database, experimental results demonstrated that features
extracted using kernel methods are more discriminative than the ones using standard methods. In addition,
compared to the LS-SVM, polynomial-based features with sMLPNN gained higher performances. Moreover,
obtained predictivity, accuracy, and area under receiver operating curve also demonstrate that kernel machines
can detect epileptic and eye states EEG signals with highest performances of 100%, 100% and 1, respectively.
Conclusion: Thus, the proposed framework can be efficient for EEG diagnosis. Overall, given the complexity and
heterogeneity of the brain, it is likely frameworks of this type that will be required to configure intelligent de­
vices for treating epilepsy and to configure eye-brain-computer interface.

1. Introduction deliver direct electrical stimulation to affected areas of the brain aim to
be the most promising. These treatments extremely depend on robust
Brain disorders are increasingly recognized as a major public health systems for seizures and eye states detection to perform effectively. In
issue. Instead of all these cerebral anomalies, epilepsy is one of the most addition, epilepsy and eye states detection remain a topic of much
common neurological diseases affecting approximately 50 million peo­ concern with a major problem which is the diagnosis accuracy [3]. Thus,
ple worldwide which is characterized by recurrent uncontrolled and efficient electroencephalogram (EEG) signals representation and feature
unprovoked seizures which have a great negative effect on the daily life extraction are important issues in developing these robust systems for
of patients [1,2]. For the diagnosis and treatment of epilepsy, many new epileptic and eye states EEG signals detection.
therapies have been investigated and implantable devices that can During these recent years, advancements and refinement of

* Corresponding author. Faculty of Science, University of Dschang-Cameroon, P.O. Box 67, Dschang.
E-mail addresses: [email protected], [email protected] (L.C. Djoufack Nkengfack).

https://doi.org/10.1016/j.imu.2021.100721
Received 29 June 2021; Received in revised form 24 August 2021; Accepted 24 August 2021
Available online 27 August 2021
2352-9148/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

automatic classification frameworks have led to the ability to detect some frameworks extend to be very complex for practical applications
epileptic seizures and eye states. An increasing interest is focused on the on EEG signals classification. On the other hand, more efforts have been
application of kernel machines for automatic EEG signals detection. In focused on the association of EEG rhythms decomposition methods and
addition, feature extraction methods are abundantly developed and features extraction methods in order to detect epileptic seizures or eye
applied in biomedical science since relevant extracted features improve states. Also, some of these frameworks do not relate the decomposed
the performances of the kernel machines. Thus, in this impetus, several rhythms to the ones defined in the literature in terms of spectral co­
frameworks have been proposed for handling epileptic and eye states efficients as shown in Ref. [17]. In short, despite several techniques
EEG signals detection, and among these frameworks, multilayer per­ applied, polynomial transforms are not yet extremely associated to
ceptron neural network (MLPNN) and support vector machines (SVM) feature extraction methods for the purpose of EEG signals classification
are two extremely used kernel machines [4–18]. At the beginning of this even if it is already shown that the physical interpretation of the spectral
last decade, Subasi et al. [4] proposed a versatile signal processing and coefficients leads to a new issue for automatic diagnosis in seizures and
analysis framework that decomposed EEG signals into the frequency eye states detection [17–19]. In this impetus, this study applies poly­
sub-bands using discrete wavelet transform (DWT) and computed a set nomial transforms namely discrete Legendre transform (DLT) and
of statistical features from these sub-bands to represent the distribution discrete Chebychev transform (DChT) to decompose EEG signals into
of wavelet coefficients. Furthermore, principal component analysis their corresponding five rhythms in terms of spectral coefficients.
(PCA), independent component analysis (ICA) and linear discriminant Moreover, knowing that EEG signals are noisy and non-stationary, linear
analysis (LDA) were used to reduce the dimension of data and extract and non-linear analysis may be effective, and the time and frequency
relevant features which were applied as inputs to a support vector ma­ distribution of the EEG signal and corresponding rhythms are analyzed
chine (SVM) with two discrete outputs and valuable performance of using measures of dispersion and measures of central tendency. Thus,
classification are obtained. This study presented and compared the the PCA, KPCA, LDA and GDA methods are applied to extract
performance of the classification process due to PCA, ICA and LDA low-dimensional features which are easily viewed and interpreted, and
methods to determine the accurate classification process such that at the generally are sufficiently representative and discriminative for epileptic
end of the study, it was revealed that the classification process including seizures and eye states detection.
LDA gained the highest performance. Hadi et al. [10] proposed a new The rest of this study is organized as follows. The materials and
scheme based on a tunable Q-factor WT (TQWT) for rhythms decom­ methods used to develop our framework are presented in Section 2. EEG
position where several statistical features are extracted and forwarded to data, polynomial-based EEG analysis, feature extraction methods, kernel
a bagging tree (BT), k-nearest neighbor (k-NN) and SVM as classifiers machines, detection tasks, metrics for performance measure and pro­
with the highest performance of 100% classification accuracy. Another posed framework are described. Section 3 presents and discusses ob­
new method called the entropy of visibility heights of hierarchical tained experimental results. This study ends with a conclusion and some
neighbors (EVHHN) is also proposed in Ref. [11] to detect seizures from concluding remarks in Section 4.
the EEG signals. First, the visibility relationships of the three nearest
neighbors are determined by a visibility criterion and the visibility 2. Materials and methods
heights of the three nearest neighbors for each data point are computed.
Next, the four different kinds of entropy associated with neighbor visi­ Several frameworks for epileptic EEG signals detection have been
bility states are calculated to characterize the EEG signals; and finally, investigated and the most promising ones can be used to design
these features are validated by LS-SVM classifiers. In their experiment, implantable devices that deliver direct electrical stimulation to affected
the normal and ictal EEG signals are classified with an accuracy of areas of the brain. This type of epilepsy treatment greatly depends on the
99.6%, meanwhile, the interictal and ictal EEG signals are distinguished robustness of the algorithm. Therefore, the motivation of this study is to
with an accuracy of 98.35%. Based on Laguerre polynomials, Ateman­ investigate and compare polynomial-based PCA, KPCA, LDA and GDA
gho and Tchiotsop [12] constructed a new basis of wavelets and applied feature extraction methods that are applied on EEG signals in order to
it in epileptic seizures detection of EEG signals with the help of PCA for enhance the epileptic and eye states EEG signals detection perfor­
dimensionality reduction, and artificial neural networks (ANN) and mances. Towards that end, Figure l presents the proposed automated
SVM for classification. This framework presented better performances in detection framework. This framework is gradually described in the
the detection of seizures than several previously proposed frameworks. following sub-sections.
Atangana et al. [15] developed an epileptic seizures detection system
where a WT-based linear discriminant feature is computed for the 2.1. EEG data
classification of EEG signals using a MLPNN classifier. This system
achieved a high accuracy of 100% in the discrimination of healthy and Most of the researchers in the field of epileptic EEG signals detection
epileptic EEG signals. For epileptic seizures and eye states detection, use the publicly available data described in Ref. [20]. The data was
Djoufack et al. [17] introduced the concept of polynomial-based collected from ten patients and consists of five sets (A-E) each containing
rhythms decomposition and analyzed EEG signals for the extraction of 100 EEG segments recorded through single channel. The first two sets
low-dimension and discriminative features using LDA. Twelve experi­ (SET A and SET B) consist of data collected from five healthy volunteers,
ments are studied and the SVM gained maximum accuracies between relaxed in an awake state with eyes opened (SET A) and closed (SET B).
96.25 and 100%. Thereafter, Djoufack et al. [18] also applied poly­ SET C and SET D segments have been measured intracranial in
nomial transforms and computed different Jacobi polynomial seizure-free intervals from five others patients in the epileptogenic zone
transforms-based measures of complexity like approximate entropy and the hippocampal formation of the opposite hemisphere of the brain,
(ApEn), sample entropy (SampEn), permutation entropy (PermEn) and respectively. SET E consists of data recorded intracranial during a
fuzzy entropy (FuzzyEn). Using these measures of complexity, they seizure. The data sets were recorded using a 128-channel amplifier
developed an automated classification system with 100% accuracy that system and standardized 10–20 electrodes placement scheme. After
can be performed to detect whether EEG signals belong to epileptic recording, they were sampled and digitized at 173.61 samples per sec­
patients in seizures or seizures-free conditions, or to normal individuals ond using 12 bits resolution. Furthermore, a bandpass filter with
with opened or closed eyes. 0.50–40 Hz (12 dB/oct) was applied. Fig. 2 below presents examples of
From the literature, it is firstly observed that most of the reported EEG signals of the data sets.
frameworks have a limited success rate and kernel machines aim to be
more applied than others classifiers machines. In addition, the ability to
extract relevant features has become increasingly important even if

2
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

2.2. Polynomial-based EEG analysis 4th qr, iqr, mad1, mad2, LL, NE, ShEn1 and ShEn2 for the delta, theta,
alpha and beta rhythms; and Max, range, mean, StD, and Var for the
2.2.1. EEG rhythms decomposition gamma rhythm.
Polynomials are extremely exploited in the literature for signal Using the samples, EEG and corresponding rhythms are analyzed in
analysis as information retrieval and compression [21–24], and feature the time domain as follow: Min, Max, range, StD, Var, 1st qr, 3rd qr, 4th
extraction for pattern classification applications [12,17–19]. This study qr, iqr, mad1, mad2, rms, AvP, norm, E, LL, and NE for the EEG signal;
applies the recently proposed scheme in Ref. [17] to decompose EEG and Min, Max, range, StD, Var, mode, 1st qr, 3rd qr, 4th qr, iqr, mad1,
signals into their delta, theta, alpha, beta and gamma rhythms before mad2, rms, AvP, norm, E, LL, and NE for the delta, theta, alpha and beta
analysis in terms of linear and non-linear parameters. This scheme is rhythms.
based on Jacobi polynomial transform and briefly described in the In addition, EEG and decomposed rhythms are analyzed using the
following and an exemplary of its application is presented by Fig. 3. energy of the approximated EEG signal (Eapp), PRD, and BP of delta to
Decompose the signal in terms of spectral coefficients using Jacobi gamma rhythms.
polynomial transforms defined by equations (1) and (2) that represent Thereafter, to have zero-score mean and unit Var before applying
the discrete Legendre polynomial transform (DLT) and the discrete features extraction methods, computed linear and non-linear parameters
Chebychev polynomial transform (DChT), respectively. are transformed. For this purpose, the entire training data are used to
⎧ ( ) determine the standardization parameter (mean) or the normalization
⎪
⎪ 2k + 1 ∑ M
Lk xj ( )
parameters (mean and Var) used to standardize or normalize the entire
⎪ αk =
⎪ [ ( )]2 s xj , k = 0, 1, ..., M − 1
⎪
⎪
⎪
⎨
M(M + 1) j=0 LM xj training and testing data according to the following equations,
(1) respectively:
⎪
⎪ ( )
⎪ ∑M
⎪
⎪
⎪
⎪ αM =
1 s xj
( ) , k=M XNew = XOld − Mean (4)
⎩ M + 1 j=0 LM xj
XOld − Mean
⎧ XNew = √̅̅̅̅̅̅̅̅ (5)
1 ∑
M +1
( ) Var
⎪
⎪ α0 = s xj
⎪
⎪
⎪ M+1
⎪
⎪
⎪ j=1 Where XOld and XNew stand to old and new feature vectors, respectively.
⎪
⎪
⎪
⎪ M +1 { [ ]}
⎪
⎪
⎪ 1 ∑ (2j − 1)
⎪ = s cos π
⎪
⎪
⎪
⎪ M + 1 j=1 2(M + 1) 2.3. Feature extraction methods
⎪
⎨
⎪ [ ] After polynomial-based EEG analysis, PCA, KPCA, LDA and GDA
⎪
⎪ 2 ∑ M +1
( ) k(2j − 1)
⎪
⎪
⎪ αk = s xj cos π feature extraction methods are applied to derive low-dimensional fea­
⎪ M + 1 j=1 2(M + 1)
⎪
⎪
⎪
⎪ tures such that much information as possible about the original data is
⎪
⎪
⎪
⎪ preserved. These extracted features can be feed as inputs of kernel ma­
⎪
⎪
⎪
⎪
⎪ 2 ∑ M +1 { [
(2j − 1)
]} [
k(2j − 1)
] chines in order to investigate epileptic EEG signals detection tasks and
⎪
⎩= s cos π *cos π , k = 1, 2, 3, ..., M detection performances can be improved.
M + 1 j=1 2(M + 1) 2(M + 1)
(2) 2.3.1. Principal component analysis (PCA)
Then, using the Fourier transform, select the spectral coefficients αk Introduced by Karl Pearson [29], PCA is a powerful unsupervised
corresponding to different rhythms and applies the inverse Jacobi linear transformation technique of extracting relevant information from
polynomial transforms described by equation (3) to provide the rhythms the original dataset and provides a roadmap for dimensionality reduc­
decomposition of the EEG signal. tion. PCA ignores class label and only find the principal components
(directions of maximum variance) in high-dimensional data that mini­
∑
M
s(x) = αk Jkα,β (x) (3) mize information loss when projecting them onto a lower-dimensional
k=0 subspace. In practice, the basic approach of the PCA is theoretically
simple to implement and is described as follows:
2.2.2. EEG analysis and standardization/normalization
After the decomposition of EEG signals using polynomial transforms, • Standardize the d-dimensional data XT = (x1 ,x2 ,…,xN ), especially if
the time and frequency distribution of the EEG signal and obtained they are measured on different scales;
rhythms are analyzed using measures of dispersion and measures of • Compute the covariance matrix Σ:
central tendency. Among the existing measures, the following are
applied [25–28]: minimum (Min), maximum (Max), range, mean, 1 ∑ N
Σ= (xi − x)T (xi − x) (6)
variance (Var), Standard deviation (StD), mode, median, quartiles (qr), N − 1 i=1
interquartile (iqr), mean absolute deviation (mad1), median absolute
deviation (mad2), norm, energy (E), average power (AvP), band power Where x is the mean vector of the data.
(BP), root mean square (rms), line length (LL), non-linear energy (NE),
impulse factor (IF), percent root square difference (PRD). • Compute and sort in decreasing order the eigenvectors and eigen­
Using the spectral coefficients of decomposition, EEG and corre­ values of the covariance matrix;
sponding rhythms are analyzed in the frequency domain as follow: StD, • Define the dxk projection matrix W = (w1 , w2 , …, wk ) of principal
Var, mad1, rms, AvP, norm, E, LL and NE for the EEG signal; Min, Max, components using the first eigenvectors that correspond to the
range, StD, Var, mode, 1st qr, 3rd qr, 4th qr, iqr, mad1, mad2, rms, AvP, largest eigenvalues;
norm, E, LL, NE and IF for the delta, theta, alpha and beta rhythms; and • In the last step, reconstruct a large fraction of the information of the
Min, Max, range, StD, Var, mode and IF for the gamma rhythm. original data X onto the new k-dimension subspace Y using:
Using the absolute values of the spectral coefficients of decomposi­ Y = X*W (7)
tion, EEG and corresponding rhythms are analyzed in the frequency
domain as follow: mean, StD, Var, iqr, mad1, LL, NE, ShEn1 and ShEn2 2.3.2. Kernel principal component analysis (KPCA)
for the EEG signal; Max, range, mean, StD, Var, median, 1st qr, 3rd qr, KPCA is an extension of the standard PCA to nonlinear problems

3
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Fig. 1. Automated epileptic and eye states EEG signals detection framework for the comparison study of PCA, KPCA, LDA and GDA feature extraction methods.

points in KPCA is much higher than in PCA and the number of principal
components that need to be estimated is also much larger. The
kernel-based learning algorithm of the KPCA is theoretically simple to
implement and is described as follows:

• Standardize the d-dimensional data XT = (x1 ,x2 ,…,xN ), especially if

they are measured on different scales;
• Define the kernel function kij = k(xi , xj );
• Construct the N-dimensional kernel matrix K (N >> d) from the d-
dimensional training data
( )
X T = (x1 , x2 , …, xN ) : K = kij i = 1, 2, ..., N (8)
j = 1, 2, ..., N

• Compute the Gram matrixK:

̃

̃ = K − 1N K − K1N + 1N K1N
K (9)
Fig. 2. Plot of some EEG signals [20].
Where 1N represents the NxN matrix with all elements equal to 1/N
[31].
using the kernel method [30]. KPCA seeks to find directions called
kernel principal components along which the classes are best separated.
• Compute and sort in decreasing order the eigenvectors and eigen­
The objective of KPCA is to extract informative features by performing
values of the Gram matrix;
dimensionality reduction. Compare to PCA, KPCA is computationally
intensive and takes a lot more time since the number of training data

4
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Fig. 3. EEG rhythms decomposition of (a) a normal EEG (Set A) and (b) a seizure EEG (Set E). Where (i) and (ii) represent the decomposed rhythms (from top to
bottom we have the delta, theta, alpha, beta and gamma rhythms) using Legendre and Chebychev polynomial transforms, respectively.

Nl
1 ∑L ∑
( l )T ( )
Sw = x − μl xil − μl (11)
N l=1 i=1 i

1 ∑L
Sb = Nl (μl − μ)T (μl − μ) (12)
N l=1

Where for a class label l, xli , Nl and μl are the row vector, the total number
of vectors and the mean vector, respectively. L, N and μ are the number
of classes, the total number of vectors of all classes, and the global mean
of all classes, respectively.

• Compute and sort in decreasing order the eigenvectors and eigen­

values of the S−w 1 Sb matrix;
• Define the dxk projection matrix W = (w1 , w2 , …, wk )of linear dis­
criminants formed by the first k-eigenvectors of S−w 1 Sb that corre­
Fig. 4. sMLPNN architecture. spond to the largest eigenvalues;
• In the last step, reconstruct a large fraction of the information of the
• Define the Nxk projection matrix W = (w1 , w2 , …, wk ) of kernel original data X onto the new k-dimension subspace Y using Equation
principal components using the first eigenvectors that correspond to (7).
the largest eigenvalues;
• Then, for each dataset Y T = (y1 , y2 , …, yL ), reconstruct a large frac­ 2.3.4. Generalized discriminant analysis (GDA)
tion of the information of the original data Y onto the new k- GDA also known as kernel LDA is an extension of the classical LDA
dimensional subspace Z using: from linear domain to non-linear domain using the kernel trick intro­
( ) duced by Baudat and Anouar [32]. GDA provides an extremely approach
Z = k yi , xj *W (10)
to extracting non-linear features since directions called generalized
discriminants are computed by mapping the training data in
2.3.3. Linear discriminant analysis (LDA) high-dimensional space using a kernel function. As LDA, GDA also tries
LDA is a powerful supervised algorithm that computes the linear to minimize the intra-classes inertia and maximize the inter-classes
discriminant by taking in consideration the scatter within-classes and inertia. Thus, our goal is to extract low-dimensional relevant informa­
the scatter between-classes [4]. In this study, the desired goal of LDA is tion using the GDA. For a given training dataXT = (x1 , x2 , …, xN ), the
to extract discriminative features by reducing the dimension of the procedure of the GDA algorithm could be summarized as follows:
original dataset. During this process, LDA tries to explicitly model the
difference between classes. The steps involved to compute the LDA are • Standardize the d-dimensional data XT , especially if they are
depicted as follows: measured separately;
• Define the kernel function kij = k(xi , xj ) = φ(xi )T φ(xj );
• Standardize the d-dimensional data XT = (x1 ,x2 ,…,xN ), especially if • Derive the NxN kernel matrix K and the NxN block-diagonal matrix
they are measured on different scales; W (N >> d) from the d-dimensional training data XT = (x1 ,x2 ,…,xN ):
• Compute the within-class (Sw ) and the between-class (Sb ) scatter ( )
matrices; K = Kpq p = 1, 2, ..., C (13)
q = 1, 2, ..., C

5
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Fig. 5. Projection of some extracted low-dimensional features using polynomial-based PCA method. Where (a) and (b) represent the DLT and DChT-based features,
and (i) to (iii) represent the C1 to C3 combinations, respectively.

Fig. 6. Projection of some extracted low-dimensional features using polynomial-based KPCA method. Where (a) and (b) represent the DLT and DChT-based features,
and (i) to (iv) represent the C1 to C4 combinations, respectively.

and (
1
)
( ) Wp = (16)
W = Wp p=1,2,...,C (14) Np Np ×Np

With Where C, Np , Nq and N stand to number of classes, number of patterns

( ) for the class p, number of patterns for the class q and total number of
Kpq = kij i = 1, 2, ..., N (15) patterns for the entire training data, respectively.
p
j = 1, 2, ..., Nq
• Centralize K by deriving the Gram matrix K:
̃
and

6
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Fig. 7. Projection of some extracted low-dimensional features using polynomial-based LDA method. Where (a) and (b) represent the DLT and DChT-based features,
and (i) to (viii) represent the C1 to C8 combinations, respectively.

̃ = K − 1N K − K1N + 1N K1N
K (17) • Then, for a given pattern x reconstruct a large fraction of its infor­
mation using the projection:
Where 1N represents the NxN matrix with all elements equal to 1/N [31]. Np Np
C ∑
∑ ( )T C ∑
∑ ( )
vT φ(x) = αpq φ xpq φ(x) = αpq k xpq , x (20)
• Using the eigenvectors decomposition, compute and sort in p=1 q=1 p=1 q=1
decreasing order the eigenvalues Γ of the Gram matrix and their
corresponding eigenvectors P;
• Compute and sort in decreasing order the eigenvalues λ and the 2.4. Detection using kernel machines
corresponding eigenvectors β of the matrix PT WP;
• Compute the eigenvectors α = P Γ − 1 β and normalize them as After the feature extraction process, two kernel learning machines
α called sMLPNN and LS-SVM are used for the detection process. These
α = √̅̅̅̅̅̅̅̅̅̅̅
̅ (18) kernel-based learning algorithms analyze and discriminate patterns by
αT K α
solving equations.

2.4.1. Simple multi-layers perceptron neural network machine (sMLPNN)

• Compute the eigenvectors v using α: The MLPNN is a popular kernel machine for data processing [33–35].
The goal is to classified different sets of EEG signals using a sMLPNN. As
shown in Fig. 4, the sMLPNN has n, three and one neurons at the input,
Np
C ∑
∑ ( )
v= αpq φ xpq (19)
p=1 q=1
hidden and output layers, respectively. Here, we have a unique hidden
layer and the input layer is not really treated as a layer of a neural
Where xpq is the pattern q of the class p. processing unit. No processing will occur in the input layer and it is only
an input vector augmented with a bias term, whose components will be
feed to the next layer. The advantage of using this sMLPNN is the rapid

7
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Fig. 8. Projection of some extracted low-dimensional features using polynomial-based GDA method. Where (a) and (b) represent the DLT and DChT-based features,
and (i) to (viii) represent the C1 to C8 combinations, respectively.

execution and generalization of the trained network, which is particu­ and

larly advantageous in the detection of EEG signals applications. η∑
The weights and bias are determined using the backpropagation al­ ωj,0 = ωj,0 + Δωj,0 = ωj,0 + δtj (22)
N
gorithm, which is based on searching a minimum mean square error
t

between the desired and obtained solutions using gradient descent [34]. where the local gradient is defined by:
The backpropagation algorithm is depicted as follows and is performed
⎧ t t( )
repeatedly until the sMLPNN solution agrees with the desired value t
⎨ ej yj 1 − yj
⎪ , if ​ j ​ is ​ a ​ neuron ​ of ​ the ​ output ​ layer
within a pre-specified tolerance. δtj = ( )∑ (23)
⎪
⎩ ytj 1 − ytj δtk ωk,j , if ​ j ​ is ​ a ​ neuron ​ of ​ the ​ hidden ​ layer
• Standardize the features X = {xt }Nt=1 on [− 1; 1] and the desired out­ k

t N
puts Y = {yt }Nt=1
̃r = on [0.05; 0.95];
{̃r }t=1 2.4.2. Least-square support vector machine (LS-SVM)
• Arbitrarily initialize weights ωj,i (weight of the link connecting the SVM originally proposed by Vapnik [36] presents many extensions
neuron j to the neuron i of the preceding layer) on [− 0.5; 0.5] and bias like the LS-SVM proposed by Suykens and Vanderwalle [37]. LS-SVM is
ωj,0 (bias of the neuron j) to one; a powerful kernel-based statistical learning methodology used for
• As long the stop criterion is not reached, repeat: non-linear classification tasks like EEG signal detection.
- Calculate the obtained outputs Y = {yt }Nt=1 by propagating the data In the LS-SVM, the problem of optimization is described as:
forwards;
- Calculate the observed errors at each neuron of the output layer etj = 1 1 ∑N
minJ(ω, b, e) = ωT ω + γ ε2 (24)
2 2 k=1 k
̃rtj − ytj ;
ω,b,e

- Adjust the weights and bias by retro-propagating the observed errors; Such that:
η∑ t [ ]
ωj,i = ωj,i + Δωj,i = ωj,i + δ (21) yk ωT ϕ(xk ) + b = 1 − εk , k = 1, 2, ..., N (25)
N t j

Where {(xk , yk )}Nk=1 , γ and εk, k=1,2,...,N stand for the training set with

8
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Table 1 following matrix formula.

Accuracy comparison of the polynomial-based PCA feature extraction method [ ][ ] [ ]
for epileptic and eye states EEG signals detection using kernel machines. 0 YT b 0
− 1 = → (27)
Y Ω+γ I α 1
Detection task Legendre-based feature Chebychev-based feature
combination combination →
with Y = [y1 , y2 , ..., yN ]T , 1 = [1, 1, ..., 1]T , ε = [ε1 , ε2 , ..., εN ]T , α =
C1 C2 C3 C1 C2 C3
[α1 , α2 , ..., αN ]T and I is the identity matrix. Mercer’s conditions for ker­
A-E sMLPNN 100 100 100 100 100 100
LS-SVM 100 61.25 100 100 100 100 nels are applied within the matrix Ω with elements given as Ωkl =
B-E sMLPNN 100 92.50 95.00 100 97.50 96.25 yk yl ϕ(xk )T ϕ(xl ) = yk yl K(xk , xl ), where K denoted the kernel function.
LS-SVM 100 77.50 98.75 100 98.75 98.75 Then, the LS-SVM classifier is expressed by:
C-E sMLPNN 100 100 100 100 100 100
[ ]
LS-SVM 100 77.50 93.75 100 98.75 93.75 [ T ] ∑N
D-E sMLPNN 97.50 97.50 97.50 97.50 100 97.50 y(x) = sign ω ϕ(xk ) + b = sign αk yk K(x, xk ) + b (28)
LS-SVM 95.00 96.25 93.75 95.00 95.00 93.75 k=1
AB - E sMLPNN 100 95.00 96.67 99.17 98.33 98.33
LS-SVM 100 79.17 99.17 100 97.50 99.17
CD - E sMLPNN 98.33 98.33 98.33 97.50 99.17 98.33 2.5. Design of detection tasks and metrics for performances measure
LS-SVM 95.83 95.83 92.50 95.83 93.33 92.50
ACD - E sMLPNN 96.88 98.75 98.75 97.50 98.13 98.75
2.5.1. Design of detection tasks
LS-SVM 96.25 94.38 96.25 96.25 91.88 96.25
ABCD - E sMLPNN 98.00 96.50 96.50 97.00 97.00 96.50
EEG signals detection using kernel machines provides an opportunity
LS-SVM 96.00 90.50 94.50 96.00 96.50 94.50 for new diagnostic and therapeutic applications like the ability to detect
AB - CD sMLPNN 93.13 80.00 82.50 90.00 88.75 81.25 whether epileptic patients are in seizure-free or seizure conditions, and
LS-SVM 91.88 80.63 86.88 91.88 83.50 87.50 whether normal patients are with opened or closed eyes. In this study,
AB - CDE sMLPNN 92.00 81.50 82.50 84.50 89.50 80.50
eleven detection tasks are elaborated and discussed using all the Bonn-
LS-SVM 88.50 82.50 89.50 88.50 83.00 89.50
A-B sMLPNN 97.50 96.25 92.50 95.00 93.75 95.00 University EEG data [20], the polynomial-based PCA, KPCA, LDA and
LS-SVM 95.00 93.75 88.75 95.00 92.50 88.75 GDA features, and the kernel-based learning algorithms. These detection
tasks are described as follows:

input xk ∈ IRn and output label yk ∈ {±1}, the regularization parameter, • A-E: Normal with eyes opened versus Epileptic with seizures;
and the errors between the obtained and the desired outputs of the LS- • B-E: Normal with eyes closed versus Epileptic with seizures;
∑
SVM with parameters ω = Nk=1 αk yk ϕ(xk ) and b, respectively. • C-E: Epileptic without seizures from the hippocampal formation
Using the Lagrangian multipliers method L to solve this dual prob­ versus Epileptic with seizures;
lem, the solution is defined by: • D-E: Epileptic without seizures from the epileptogenic zone versus
∑
N Epileptic with seizures;
{ [ ] }
L(ω, b, ε; α) = J(ω, b, ε) − αk yk ωT ϕ(xk ) + b − 1 + εk (26) • AB-E: Normal versus Epileptic with seizures;
k=1 • CD-E: Epileptic without seizures versus Epileptic with seizures;
• ACD-E: Normal with eyes opened and Epileptic without seizures
And the set of corresponding linear equations can be defined as the
versus Epileptic with seizures;

Fig. 9. ROC curves with corresponding AUC values for the highest detection performances of different detection tasks. For each detection task, curves in red and blue
correspond to the training ROC and testing ROC, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web
version of this article.)

9
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Table 2
Accuracy comparison of the polynomial-based KPCA feature extraction method for epileptic and eye states EEG signals detection using kernel machines.
Detection task Legendre-based feature combination Chebychev-based feature combination

C1 C2 C3 C4 C1 C2 C3 C4

A-E sMLPNN 100 100 100 100 100 100 100 100
LS-SVM 100 100 100 100 100 100 100 100
B-E sMLPNN 100 100 98.75 100 98.75 100 98.75 98.75
LS-SVM 100 100 100 100 100 100 100 98.75
C-E sMLPNN 100 100 98.75 100 100 100 98.75 100
LS-SVM 100 100 100 100 100 100 100 100
D-E sMLPNN 97.50 96.25 97.50 97.50 98.75 97.50 95.00 97.50
LS-SVM 97.50 96.25 97.50 97.50 98.75 96.25 96.25 97.50
AB - E sMLPNN 100 100 99.17 100 99.17 100 99.17 99.17
LS-SVM 100 100 100 100 100 100 100 98.33
CD - E sMLPNN 98.33 97.50 97.50 98.33 98.33 97.50 96.67 98.33
LS-SVM 98.33 97.50 97.50 98.33 98.33 97.50 96.67 98.33
ACD - E sMLPNN 98.75 98.13 98.13 99.38 99.38 98.13 97.50 99.38
LS-SVM 98.75 98.13 98.13 98.75 98.75 98.13 97.50 98.75
ABCD - E sMLPNN 98.50 98.50 98.50 97.50 97.50 98.00 97.50 97.50
LS-SVM 98.50 98.50 98.00 97.50 97.50 98.50 97.50 97.00
AB - CD sMLPNN 90.63 96.25 96.25 91.25 94.38 96.25 95.63 95.00
LS-SVM 74.38 74.38 76.88 63.75 87.50 78.76 75.63 69.38
AB - CDE sMLPNN 92.50 96.50 98.00 92.50 95.00 95.50 95.00 95.00
LS-SVM 81.50 76.00 85.50 79.00 92.50 67.00 92.50 75.50
A-B sMLPNN 96.25 86.25 95.00 97.50 92.50 88.75 96.25 98.75
LS-SVM 86.25 66.25 95.00 88.75 86.25 61.25 95.00 81.25

• ABCD-E: Normal and Epileptic without seizures versus Epileptic with against the x-axis. In practice, a single measure called area under the
seizures; ROC curve (AUC) is used to assess the performance or discriminative
• AB-CD: Normal versus Epileptic without seizures; ability of a framework, or to compare different classifiers, indepen­
• AB-CDE: Normal versus Epileptic; dent of a specific threshold. The AUC performs well as a general
• A-B: Normal with eyes opened versus Normal with eyes closed. measure so that the higher AUC shows the better diagnostic system.

2.5.2. Metrics for performances measure Where TN , FP , TP and FN stand to the total number of true negative
To assess the performances of the proposed framework, the 10-fold (normal or seizure-free), false positive, true positive (epileptic) and false
cross-validation technique is applied. The data is firstly partitioned negative, respectively.
into training and testing data. Secondly, the training data is randomly
partitioned into 10 subsets of the same size and with the same number of
EEG signals. One subset is taken as the validation subset, whereas the 2.6. Proposed detection framework description
remainder is used as the training subset. This process is repeated 10
times, such that every subset is used as a validation subset exactly once. A novel full EEG signals detection framework for epileptic seizures
The average parameter of all training experiments is used as the optimal and eye states detection which applied the recently proposed alternative
parameter of the kernel machine. Then, the constructed kernel machines approach for EEG rhythms decomposition is proposed and presented in
are retained and the effectiveness of the proposed framework is assessed Fig. 1. The JPTs is firstly applied to decompose the entire Bonn-
using the testing data according to five metrics for performances mea­ University EEG database into their five rhythms. Thereafter, EEG and
sure. These metrics namely specificity (Spe), sensitivity (Sen), pre­ rhythms are analyzed using measures of dispersion and central ten­
dictivity (Pre), accuracy (Acc) and receiver operating curve (ROC) are dency. These measures are standardized or normalized and different
defined in the following [38–40]. feature extraction methods can be applied. For feature extraction, the
principal components (PC), kernel PC (KPC), linear discriminants (LD)
Spe(%) = 100 ×
TN
(29) and generalized discriminants (GD) corresponding to the application of
T N + FP the PCA, KPCA, LDA and GDA methods, respectively, are computed
using the first 50% of SETs A to E. Then, for each detection task, the
TP
Sen(%) = 100 × (30) corresponding measures are projected using these computed PC, KPC,
TP + FN
LD and GD such that informative and discriminative features are
TP extracted and feed as inputs of two kernel machines namely sMLPNN
Pre(%) = 100 × (31) and LS-SVM. Therefore, to carry out each detection task, the data is
TP + FP
partitioned such that the first 60% of each set of EEG signals constitute
TP + TN the training set while the remaining 40% constitute the testing set.
Acc(%) = 100 × (32)
TP + FN + TN + FP It is wise to remember that, while taking into account the literature,
the configuration parameters of the KPCA and the GDA were selected
empirically after a multitude of tests to extract the most prominent and
convincing features for discrimination as well adequate as possible.
• The ROC curve makes use of the fact that many classification Therefore, these tests made it possible to set the parameters of the KPCA
frameworks generate probabilities of assigning an input data sample and GDA such that the hyperbolic tangent kernel with delta=5 and the
to each possible output label. It represents a graphical way to classify Gaussian kernel with sigma=10 are used for the KPCA and GDA,
the performance and globally visualize the behavior of a binary respectively. At the last stage, the sMLPNN with the sigmoidal activation
diagnostic system. The sensitivity or true positive rate is plotted 1
function K(xk ) = 1+exp(− xk ), and the LS-SVM with Gaussian radial basis
against the y-axis and the 1-specificity or false positive rate is plotted ( )
function (RBF) kernel K(x, xk ) = exp − 21σ2 ‖xk − x‖ are applied as

10
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

kernel machines such that a decision can be taken as seizure or non-

66.253
seizure, or as normal with opened or closed eyes.

97.50
96.25

95.00
96.25
92.50
98.33
95.83
96.67
92.50
96.88
94.38
97.50
94.50
71.88

73.00
68.50
85.00
86.25
100
100
100
C8 Several existing methods are combined to design our framework
such that the detection of epileptic seizures and eye states can be opti­
mized. The novelty of this work is the combination of the novel alter­

98.75
98.75

99.17
99.17
99.38
98.75
98.75

99.17
99.17

98.13
97.50
98.00
92.50
88.75
93.00
88.00
96.25
91.25
native approach of EEG rhythms decomposition using JPTs to PCA,

100
100
100
100
C7

KPCA, LDA and GDA feature extraction methods to extract prominent

and discriminative low-dimension feature vectors instead of high
98.75
98.75
98.75
98.75
91.25
93.75
90.00
99.17
95.83
95.83
92.50
96.88
93.75
97.00
94.50
81.25
81.88
80.00
73.00
90.00
83.75
dimension features such that the classification performances can be
100
C6

significantly improved for diagnosis using EEG signals. This also helps to
establish correlations of spectral coefficients to some pathologies. In
addition, to the best of our knowledge, this work is one of the first that
98.75
97.50
98.75
98.75
93.75
95.00
99.17
99.17
95.83
96.67
96.88
97.50
97.50
98.00
68.75
70.63
76.50
75.00
83.75
73.75
100
100

applied the novel proposed JPTs method for rhythms decomposition

C5

before applying different feature extraction methods. Overall, the pro­

posed framework can be useful to clinicians in providing an accurate and
98.75

96.25
96.25

95.83
95.83
96.88
97.50
98.00
97.50
88.13
78.75
91.00
72.50
78.75
71.25
100
100

100

100
100

objective scheme in epileptic seizures and eye states detection to

100
100
C4

improve the living conditions of patients.

Chebychev-based feature combination

98.75
98.75
96.25
92.50

95.83
92.50
97.50
94.38
97.50
96.50
90.63
90.63
90.00
84.50
76.25
75.00
3. Experimental results and discussion
100
100

100
100
100
100
C3
Accuracy comparison of the polynomial-based LDA feature extraction method for epileptic and eye states EEG signals detection using kernel machines.

3.1. Experimental results

97.5096.25

The polynomial-based PCA, KPCA, LDA and GDA feature extraction

98.75
98.75

95.00
95.00
99.17
99.17
96.67
96.67

97.50
96.00
89.38
89.38
90.50
87.50
92.50
87.50

methods presented in the previous section are applied to the entire EEG
100
100
100
100
C2

database of 500 signals. In this study, the first 50% of each EEG set are
used to compute the PC, KPC, LD and GD related to the PCA, KPCA, LDA
98.75
98.75
98.75
98.75
96.25
96.25
99.17
97.50
97.50
96.67
96.38
96.25
97.50
96.50
90.00
89.38
92.00
86.00
83.75
81.25

and GDA feature extraction methods. Thereafter, these computed prin­

100
100
C1

cipal components and discriminants are used to extract low-dimensional

features for the entire EEG database. Therefore, according to the poly­
nomial methods used for rhythms decomposition (DLT or DChT), the
97.50
96.25
98.75
95.00
93.75
92.50
98.33
95.00
95.83
91.67
97.50
95.00
97.50
94.50
70.63
65.63
74.00
71.00
87.50
86.25
100
100
C8

feature extraction method, and the set of computed measures of

dispersion and central tendency, different combinations of low-
dimensional features that are more visually representative and infor­
98.75
98.75

99.17
99.17
99.38
98.75
98.75

99.17
99.17

98.13
98.00
98.00
92.50
87.50
92.50
86.00
95.00
87.50
100
100
100
100

mative for classes discrimination can be obtained. Then, some of these

C7

extracted low-dimensional features are projected and presented in Fig. 5

to Fig. 8.
98.50
97.50

96.25

96.25
97.50
95.00

96.67
98.33
95.83
98.75
95.63

96.00
83.13
73.13
83.50
71.50
92.50
95.00

For Fig. 5, computed measures of dispersion and central tendency are

100

100

100
100
C6

normalized before applying the PCA feature extraction method such that
the first combination (C1) is obtained using all measures computed on
spectral coefficients of decomposition of all rhythms and EEG; while the
98.50
97.50

98.75
97.50
96.25

98.33
98.33
96.67
98.75
97.50

98.00
75.63
71.88
79.00
73.50
86.25
86.25
100

100

100
100
100

second combination (C2) and the third combination (C3) are obtained
C5

using only measures computed on spectral coefficients of decomposition

and on samples of EEG signal, respectively.
95.00
96.25

96.67
96.67
97.50
97.50
98.00
98.00
90.63
85.63
92.00
88.00
80.00
80.00
100
100
100
100

100
100
100
100

For Fig. 6, computed measures of dispersion and central tendency are

C4

normalized before applying the KPCA feature extraction method. Based

on spectral coefficients of decomposition two 3-dimensional feature
95.00
96.25

96.67
96.67
96.88
97.50
97.50
98.00
95.00
93.75
94.50
92.00
81.25
81.25

vectors are computed such that the first 3-dimensional combination (C1)
100
100
100
100

100
100
100
100
C3

is obtained using the first KPCs computed on alpha rhythm, EEG signal,
Legendre-based feature combination

and association of all rhythms and EEG; and the second 3-dimensional
combination (C2) is obtained using the first KPCs computed on delta
96.8897.50

rhythm, gamma rhythm, and association of all rhythms and EEG. On the
97.50
97.50
97.50

98.33
97.50
98.75
95.63
98.00
96.50

96.50
85.50
96.25

other hand, based on absolute values of spectral coefficients of decom­

100
100
100
100

100
100
100
100
C2

position, the third combination (C3) is obtained using the first KPCs
computed on delta rhythm and EEG signal; and the fourth combination
88.7590.00

(C4) is obtained using the first KPCs computed on alpha rhythm, EEG
signal, and association of all rhythms and EEG.
98.75
98.75

99.17

99.38

98.13
98.13
97.00
98.33

97.50
98.00
97.50

91.00
100
100
100
100

100
100
100
100

For Fig. 7, computed measures of dispersion and central tendency are

C1

standardized or normalized before applying the LDA feature extraction

method. Based on spectral coefficients of decomposition, computed
sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN

sMLPNN
LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

LS-SVM

measures are standardized and five combinations are obtained. The first
2-dimensional combination (C1) is obtained using the second LDs
computed on delta and alpha rhythms; the second combination (C2) is
Detection task

obtained using the second LDs on delta rhythm and EEG signal; the third
ABCD - E

AB - CDE

combination (C3) is obtained using the first LDs on the delta, alpha and
AB - CD
ACD - E
Table 3

CD - E
AB - E
D-E

A-B
A-E

gamma rhythms; and the fourth combination (C4) is obtained using the
B-E

C-E

first LDs on delta and gamma rhythms. In addition, fourth other

11
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Table 4
Accuracy comparison of the polynomial-based GDA feature extraction method for epileptic and eye states EEG signals detection using kernel machines.
Detection task Legendre-based feature combination Chebychev-based feature combination

C1 C2 C3 C4 C5 C6 C7 C8 C1 C2 C3 C4 C5 C6 C7 C8

A-E sMLPNN 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
LS-SVM 100 100 100 100 100 100 100 97.50 100 100 100 100 100 100 100 100
B-E sMLPNN 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
LS-SVM 100 100 100 100 100 100 100 97.50 100 100 100 100 100 100 100 100
C-E sMLPNN 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
LS-SVM 100 100 100 100 100 100 100 98.75 100 100 100 100 100 100 100 100
D-E sMLPNN 98.75 100 100 98.75 100 100 100 100 98.75 100 98.75 96.25 98.75 98.75 96.25 97.50
LS-SVM 97.50 100 100 98.75 100 100 100 97.50 98.75 100 98.75 96.25 97.50 98.75 95.00 97.50
AB - E sMLPNN 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100
LS-SVM 100 100 100 100 100 100 100 98.33 100 100 100 100 100 100 100 100
CD - E sMLPNN 99.17 100 100 99.17 100 100 100 100 99.17 100 99.17 98.33 98.33 98.33 96.67 98.33
LS-SVM 98.33 100 100 99.17 100 100 100 99.17 99.17 100 99.17 97.50 96.67 96.67 95.00 98.33
ACD - E sMLPNN 99.38 100 100 99.38 100 100 100 100 99.38 100 99.38 98.13 98.75 98.13 98.13 98.75
LS-SVM 98.75 100 100 99.38 100 100 100 99.38 99.38 100 99.38 98.13 96.88 97.50 96.25 98.75
ABCD - E sMLPNN 99.50 100 100 99.50 100 100 100 100 99.50 100 99.00 98.50 98.50 98.50 98.50 99.50
LS-SVM 99.00 100 100 99.50 100 100 100 99.00 99.50 100 99.00 98.50 97.50 98.00 97.50 99.00
AB - CD sMLPNN 83.13 70.00 78.75 80.50 81.25 71.25 72.51 85.63 90.63 71.25 84.38 85.00 88.75 71.25 75.63 86.88
LS-SVM 75.63 68.13 75.63 83.13 79.38 71.25 76.25 85.00 86.88 68.75 79.38 84.38 83.75 69.38 73.75 87.50
AB - CDE sMLPNN 83.50 75.00 83.00 88.00 84.00 77.50 79.00 87.00 92.50 76.50 87.50 85.00 88.50 77.00 81.00 89.00
LS-SVM 84.00 71.50 80.50 83.50 85.00 75.50 78.50 86.50 88.50 76.00 85.00 89.00 87.00 75.00 77.50 91.00
A-B sMLPNN 91.25 82.50 85.00 83.75 86.25 75.00 77.50 86.25 96.25 85.00 86.25 81.25 81.25 75.00 77.50 90.00
LS-SVM 87.50 85.00 81.25 86.25 81.25 75.00 80.00 85.00 92.50 83.75 85.00 85.00 80.00 75.00 75.00 86.25

combinations are obtained based on samples of decomposed rhythms the training set and the remaining 40% of each EEG signals set form the
and EEG signals. These combinations are stated as follows: the fifth testing set. For the implementation, the MLPNN toolbox proposed in
combination (C5) is obtained using the first LDs computed on delta MATLAB R2013a and the LS-SVM toolbox proposed in Ref. [41] are used
rhythm, theta rhythm and EEG signal; the sixth combination (C6) is in this study. The 10-fold cross-validation technique is associated with
obtained using the second LDs computed on theta rhythm, gamma the training set to determine the optimal sMLPNN and LS-SVM kernel
rhythm and EEG signal after standardization of computed measures; the machines. Therefore, the constructed optimal kernel machines are tested
seventh combination (C7) is obtained using the first and sixth LDs using the testing set and their epilepsy and eye states detection accu­
computed on EEG signal after standardization of computed measures; racies are summarized and compared in Table 1 to Table 4, where bold
and the eighth combination (C8) is obtained using the second, third and indicates the highest performance for each case. Then, for each task,
fourth LDs computed on EEG signal after normalization of computed Table 5 summarizes the detection performances of the best epilepsy and
measures. eye states detection systems, and their corresponding ROC with area
For Fig. 8, computed measures of dispersion and central tendency are under curve are plotted in Fig. 9.
normalized before applying the GDA feature extraction method. Based
on spectral coefficients of decomposition, a first 3-dimensional combi­
3.2. Discussion
nation (C1) is obtained using the first GDs computed on alpha rhythm,
EEG signal, and association of all rhythms and EEG. In addition, seven
From the comparisons presented in Figs. 5–8, it is clearly shown that
other feature vectors are computed using the association of all computed
for the 2-dimensional and 3-dimensional extracted features, epileptic
measures of dispersion and central tendency in frequency and time do­
with seizures EEG signals are more distant to normal and epileptic
mains. These computed feature vectors are stated as follows: the second
without seizures EEG signals. In some cases, the overlaps of the
combination (C2) is obtained using the first GDs computed on alpha
extracted features are significantly low. In addition, these comparisons
rhythm, and the association of all rhythms and EEG; the third combi­
demonstrate that the KPCA method presents high separation ability than
nation (C3) is obtained using the first GDs computed on beta rhythm,
the PCA method. Compare to the LDA method, these comparisons also
and the association of all rhythms and EEG; the fourth combination (C4)
demonstrate the high ability of the GDA method to maximize the
is obtained using the first GDs computed on beta rhythm, EEG signal,
between-classes distance and to minimize the within-classes distance
and association of all rhythms and EEG signal; the fifth combination (C5)
than the LDA. In general, it is shown that kernel-based feature extraction
is obtained using the first GDs computed on beta rhythm, gamma
algorithms provide more informative and discriminative low-
rhythm, and association of all rhythms and EEG; the sixth combination
dimensional features than standard algorithms. These comparisons
(C6) is obtained using the first GDs computed on gamma rhythm, and
also present different epileptic seizures and eye states detection tasks
the association of all rhythms and EEG; the seventh combination (C7) is
and permit to visualize and demonstrate that the extracted features are
obtained using the first GDs computed on gamma rhythm, EEG signal,
able to detect whether EEG signals belong to epileptic patients with
and association of all rhythms and EEG signal; and the eighth combi­
seizures or without seizures, or to normal patients with closed or opened
nation (C8) is obtained using the first GDs computed on theta rhythm,
eyes.
beta rhythm, and association of all rhythms and EEG.
Keeping in mind the outcome of the above comparisons on the
From the comparisons presented in Figs. 5–8, it is observed that all
database, it is clear that extracted features can be easily fed as inputs of
the extracted features by themselves can be sufficiently used to detect
kernel machines and high performances can be obtained. For this pur­
epileptic and eye states EEG signals. Keeping in mind the outcome of the
pose, the sMLPNN and LS-SVM kernel machines are used to examine and
above comparisons on the database, it is clear that extracted features can
discuss previously defined EEG signals detection tasks. Table 1 shows
be easily fed as inputs of kernel machines and high performances can be
that the association of polynomial transforms to PCA for the analysis and
obtained. The sMLPNN and LS-SVM kernel machines are used to
detection of EEG signals gives the highest performance of 100% for the
examine and discuss previously defined epilepsy and eye states detec­
A-E, B-E, C-E and AB-E detection tasks, and higher performances be­
tion tasks. For different tasks, the first 60% of each EEG signals set form
tween 92.00% and 99.17% for the remaining seven detection tasks. In

12
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Table 5
Summary of the comparative study of the polynomial-based PCA, KPCA, LDA and GDA feature extraction methods for epileptic and eye states EEG signals detection
using kernel machines.
Detection Automated detection system Spe Sen Pre Acc
task (%) (%) (%) (%)
PCA KPCA LDA GDA

A-E JPTs – C1 to C3 – JPTs – C1 to C4 – JPTs – C1 to C8 – sMLPNN JPTs – C1 to C8 – sMLPNN 100 100 100 100
sMLPNN sMLPNN JPTs – C1 to C5, C7, C8 – LS- DLT – C1 to C7 – LS-SVM
DLT – C1, C3 – LS-SVM JPTs – C1 to C4 – LS- SVM DChT – C1 to C8 – LS-SVM
DChT – C1 to C3 – LS- SVM
SVM
B-E JPTs – C1 – sMLPNN DLT – C1, C2, C4 – DLT – C1 to C6 – sMLPNN JPTs – C1 to C8 – sMLPNN 100 100 100 100
JPTs – C1 – LS-SVM sMLPNN DLT – C1 to C4 – LS-SVM DLT – C1 to C7 – LS-SVM
DChT – C2 – sMLPNN DChT – C3, C4 – sMLPNN, LS- DChT – C1 to C8 – LS-SVM
DLT – C1 to C4 – LS-SVM SVM
DChT – C1 to C3 – LS-
SVM
C-E JPTs – C1 to C3 – JPTs – C1, C2, C3 – DLT – C1 to C7 – sMLPNN JPTs – C1 to C8 – sMLPNN 100 100 100 100
sMLPNN sMLPNN DChT – C2, C7, C8 – sMLPNN DLT – C1 to C7 – LS-SVM
JPTs – C1 – LS-SVM JPTs – C1 to C4 – LS- DLT – C1 to C4, C7 – LS-SVM DChT – C1 to C8 – LS-SVM
SVM DChT – C2, C4, C7 – LS-SVM
D-E DChT – C2 – sMLPNN – – DLT – C2, C3, C5 to C8 – 100 100 100 100
sMLPNN
DLT – C2, C3, C5 to C7 – LS-
SVM
DChT – C2 – sMLPNN, LS-
SVM
AB - E DLT – C1 – sMLPNN DLT – C1, C2, C4 – DLT – C1 to C6 – sMLPNN JPTs – C1 to C8 – sMLPNN 100 100 100 100
JPTs – C1 – LS-SVM sMLPNN DLT – C1 to C4 – LS-SVM DLT – C1 to C7 – LS-SVM
DChT – C2 – sMLPNN DChT – C3, C4 – sMLPNN, LS- DChT – C1 to C8 – LS-SVM
DLT – C1 to C4 – LS-SVM SVM
DChT – C1 to C3 – LS-
SVM
CD - E – – – DLT – C2, C3, C5 to C8 – 100 100 100 100
sMLPNN
DLT – C2, C3, C5 to C7 – LS-
SVM
DChT – C2 – sMLPNN, LS-
SVM
ACD - E – – – DLT – C2, C3, C5 to C8 – 100 100 100 100
sMLPNN
DLT – C2, C3, C5 to C7 – LS-
SVM
DChT – C2 – sMLPNN, LS-
SVM
ABCD - E – – – DLT – C2, C3, C5 to C8 – 100 100 100 100
sMLPNN
DLT – C2, C3, C5 to C7 – LS-
SVM
DChT – C2 – sMLPNN, LS-
SVM
AB - CD – – DLT – C1 – sMLPNN, LS-SVM – 100 96.25 100 98.13
AB - CDE – DLT – C3 – sMLPNN – – 100 96.67 100 98.00
A-B – DChT – C4 – sMLPNN – – 100 97.50 100 98.75

line with this, Table 2 demonstrates that the association of polynomial epileptic with seizures (ACD-E, ABCD-E) detection tasks, kernel ma­
transforms to the kernel extension of PCA namely KPCA extends to chines with polynomial-based low-dimensional GDA features fully
ameliorate previously obtained results such that higher performances discriminate classes and detect epileptic EEG signals with the highest
between 96.25% and 99.38% are obtained for the same remaining seven detection accuracy of 100%. On the other hand, the polynomial-based
detection tasks. As PCA and KPCA, Table 3 shows that the association of low-dimensional PCA, KPCA and LDA fully discriminate and detect
polynomial transforms to LDA for the analysis and detection of EEG epileptic seizures only for four of the same seven tasks. In addition, the
signals also give the highest performance of 100% for the A-E, B-E, C-E computed polynomial-based low-dimensional LDA feature as inputs of
and AB-E detection tasks, and higher performances between 97.00% and the kernel machines are able to discriminate epileptic without seizures
99.38% for the remaining seven detection tasks. On the other hand, EEG signals (CD) from normal ones (AB) with a 98.13% accuracy. Also,
Table 4 demonstrates that the association of polynomial transforms to results show that the polynomial-based low-dimensional KPCA features
the kernel extension of LDA namely GDA extends to ameliorate all as inputs of the sMLPNN are able to discriminate epileptic EEG signals
previously obtained results such that the highest performance of 100% is (CDE) from normal ones (AB) with a maximum detection accuracy of
obtained for the first eight detection tasks (A-E, B-E, C-E, D-E, AB-E, CD- 98.00%; and to detect whether EEG signals belong to normal patients
E, ACD-E, ABCD-E) and higher performances between 90.63% and with opened (A) or closed (B) eyes with maximum accuracy of 98.75%.
96.25% are obtained for the remaining three detection tasks. In sum, the Furthermore, in regard to Table 5 and for different detection tasks,
highest detection performances are presented in Table 5, and it is easily the training and testing ROC curves with corresponding AUC for the
viewed that for normal versus epileptic with seizures (A-E, B-E and AB- highest detection performances are presented in Fig. 9. The Sen is
E), epileptic without seizures versus epileptic with seizures (C-E, D-E, plotted as a function of the 1-Spe for the training and testing data. In
CD-E), and normal associated to epileptic without seizures versus general, as shown in Fig. 9, the ROC curve with corresponding AUC

13
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

Table 6
Comparison of the detection accuracies for different detection task using the same Bonn University EEG data.
Authors Frameworks Acc (%) of detection tasks

A-E B-E C-E D-E AB-E CD-E ACD- ABCD- AB- AB- A-B
E E CDE CD

Ocak, 2009 [42] DWT-based ApEn - ANN 99.60 - - - - - 96.65 - - - -

Yuedong and SampEn - BPNN/ELM - - - - - - - 95.67 - - -
Pietro, 2010
[43]
Nicolaou et al., PermEn - SVM 93.55 82.88 88.00 79.94 - - - - - - -
2012 [5]
Yatindra et al., DWT - ApEn - ANN 100 92.50 100 95.00 – – 94.00 94.00 - - -
2014 [6]
Jie Xiang et al., FuzzyEn - SVM – – 100 100 - - - - - - -
2015 [44]
Mohd Zuhair DWT - ApEn - ANN/LS-SVM 100 100 100 100 – – 100 99.50 - - -
et al., 2017
[7]
Lina Wang et al., DWT - Multi-domain features extraction - – - - - - - 99.25 - - -
2017 [45] Nonlinear analysis
Li et al., 2018 FuzzyEn/DistEn - QD classifier – – - - - 91.00 - - 93.00 - -
[8]
Gupta and FBSE based rhythms - WMRPE - LS-SVM/ 99.50 99.50 99.50 97.50 – 99.00 - 98.60 - - -
Pachori, 2019 Regression
[46]
Gurwinder et al., EMD - Fusion of non-linear features (NLF) and 100 100 100 100 – - - 100 - - -
2019 [47] spike-based features (SBF) - grasshopper
optimization algorithm (GOA) - Ensemble
classifier
Ming et al., 2019 EVHHN – LS-SVM 100 98.50 99.67 98.35 – – - - -
[11]
Khademul et al., Multiband implementation of three spike-related 100 100 99.90 98.60 100 99.49 99.34 99.39 - - -
2020 [48] features (SrF) - Four entropy features with GED-
based feature selection - KNN classifier
Mingyang et al., Time-frequency image-based STFT - Block texture 100 100 100 100 – – - 99.60 99.80 - -
2020 [49] features - SVM
Jiang et al., CEEMD - XGBoost 100 100 99.50 100 100 99.33 99.75 99.60
2020 [50]
Mustafa et al., STFT-based alpha band decomposition - 98.00 96.00 98.00 95.50 96.67 96.33 - 97.40 88.40 86.00 -
2020 [51] Statistical features - Machine learning
Wei et al., 2020 One-dimensional deep neural network 99.52 99.11 98.02 97.63 99.38 98.03 – 98.76 - - -
[14]
This work Polynomial-based PCA, KPCA, LDA and GDA 100 100 100 100 100 100 100 100 98.13 98.00 98.75
feature extraction methods – Kernel machines

Bold indicates the highest accuracy for each detection task.

summarizes the performances of each system and demonstrates the the best of our knowledge, this work is the first that for the epileptic and
highest discriminative ability of the proposed framework. In order way, eye states detection purpose discuss and compare the application of
from the AUC = 1 it is related that perfect discrimination of classes can PCA, KPCA, LDA and GDA features extraction methods. Then, in com­
be achieved using the corresponding system. In short, obtained detec­ parison with other previously proposed frameworks in the literature, it
tion performances are higher since different curves pass through the is related that relevant information are extracted from raw EEG signals
upper left corner of the ROC space and obtained AUC between 0.958 and and our proposed framework exhibits potential desirable and promising
1 are still proportional to the detection accuracies. These training and applications for medical treatment as implantable devices to intervene
testing ROC curves with corresponding higher AUC also demonstrate at right time.
that the constructed classifiers perform well on training and testing data With regard to metrics of performances measure, this comparative
such that there is hardly any over-fitting produced and a randomly work firstly revealed that the association of polynomial transforms to
epileptic seizure or eye states can be correctly detected with higher feature extraction methods improve the performance of the kernel ma­
probability. chines. Secondly, the Legendre-based combinations extend to be more
Given the complexity and heterogeneity of the brain, linear and non- discriminative than the Chebychev ones; and the polynomial-based GDA
linear feature extraction methods that extend to provide relevant and feature extraction method provides the most powerful information for
powerful information for EEG signals detection have been abundantly EEG signals discrimination compare to PCA, KPCA, LDA and GDA
applied during these last decades. In line with this, several frameworks methods. In addition, all systems tested in this work had a low level of
that use the same Bonn-University EEG data [20] have been proposed complexity (low-dimensional feature vector and a few hidden nodes in
and amount of them can be summarized and compared to our proposed the case of sMLPNN) and requested low computational time. On the
framework as shown by Table 6. other hand, the assertion that a neural network with one hidden layer
From the supervision of Table 6, it is firstly shown that most of the generally produces excellent results is confirmed since sMLPNN ob­
previously proposed works don’t examine all of the designed eleven tained better performances compared to LS-SVM during the detection.
detection tasks. In addition, sMLPNN and LS-SVM are used instead of Overall, it is found that this work presents a novel framework using
complex classifiers, and similar or higher results are obtained with our polynomials-based PCA, KPCA, LDA and GDA feature extraction
proposed framework even if previous works extend to associate different methods and kernel machines that is used to be more efficient and ac­
feature extraction methods to different classifiers. Furthermore and to curate for epileptic seizures and eye states detection. In sum, it is likely

14
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

methods of this type that will be appropriate and required to configure [10] Al Ghayab Hadi Ratham, Li Yan, Siuly S, Abdulla Shahab. A feature extraction
technique Based on tunable Q-factor wavelet Transform for brain signal
intelligent devices for treating epilepsy to each individual’s neuro­
classification. J Neurosci Methods 2019;312:43–52.
physiology prior to clinical operation. [11] Zeng Ming, Zhao Chun-Yu, Meng Qing-Hao. Detecting seizures from EEG signals
using the entropy of visibility heights of hierarchical neighbors, vol. 7; 2019.
https://doi.org/10.1109/ACCESS.2019.2890895.
4. Conclusion [12] Bruno Peachap Atemangho, Tchiotsop Daniel. Epileptic seizures detection based on
some new Laguerre polynomial wavelets, artificial neural networks and support
The ability to detect or predict the onset of seizures provides an vector machines. Informatics in Medicine Unlocked 2019;16. https://doi.org/
10.1016/j.imu.2019.100209.
opportunity for new diagnostic and therapeutic applications that could [13] Romain Atangana, Daniel Tchiotsop, Godpromesse Kenne, Djoufack
revolutionize the medical and surgical management of epilepsy. The Nkengfack Laurent Chanel. Suitable mother wavelet selection for EEG signals
kernel machines with polynomial-based PCA, KPCA, LDA and GDA analysis: frequency bands decomposition and discriminative feature selection.
Signal & Image Processing: Int J February 2020;11(No. 1):33–49. https://doi.org/
features for EEG signals detection tasks are discussed in this study. The 10.5121/sipij.2020.11104.
proposed framework firstly processes EEG signals to decompose them [14] Zhao Wei, Zhao Wenbing, Wang Wenfeng, Jiang Xiaolu, Zhang Xiaodong,
into their five bands and analyze them as a pattern vector capable to Peng Yonghong, Zhang Baocan, Zhang Guokai. A novel deep neural network for
robust detection of seizures using EEG signals. Computational and Mathematical
represent their changes in a compact way. Furthermore, the PCA, KPCA,
Methods in Medicine 2020;2020. https://doi.org/10.1155/2020/9689821.
LDA and GDA methods are applied to extract low-dimensional features. [15] Romain Atangana, Daniel Tchiotsop, Godpromesse Kenne, Djoufack
These extracted features are then used as a benchmark for epileptic and Nkengfack Laurent Chanel. EEG signal classification using LDA and MLP classifier.
Health Informatics - An International Journal (HIIJ) February 2020;9(No. 1):
eye states EEG signals detection using the sMLPNN and LS-SVM kernel
14–32. https://doi.org/10.5121/hiij.2020.9102.
machines. Towards that end, the proposed framework gained a [16] Siddiqui Mohammad Khubeb, Morales-Menendez Ruben, Huang Xiaodi,
maximum performance between 98.13 and 100% in detecting epileptic Hussain Nasir. A review of epileptic seizure detection using machine learning
and eye states EEG signals. In general, this study deeply investigates the classifers. Brain Informatics 2020;7(No. 5). https://doi.org/10.1186/s40708-020-
00105-1.
use of feature extraction methods for EEG signals detection using kernel [17] Djoufack Nkengfack Laurent Chanel, Daniel Tchiotsop, Romain Atangana,
machines. Then, obtained experimental results indicated that our pro­ Valérie Louis-Door, Wolf Didier. EEG signals analysis for epileptic seizures
posed framework is useful for clinicians in detecting whether EEG sig­ detection using polynomial transforms, linear discriminant analysis and support
vector machines. Biomed Signal Process Contr 2020;62. https://doi.org/10.1016/j.
nals belong to epileptic patients with seizures or without seizures, or to bspc.2020.102141.
normal patients with closed or opened eyes. Overall, it is likely methods [18] Djoufack Nkengfack Laurent Chanel, Daniel Tchiotsop, Romain Atangana,
of this type that will be required to configure intelligent devices for Valérie Louis-Door, Wolf Didier. Classification of EEG signals for epileptic seizures
detection and eye states identification using Jacobi polynomial transforms-based
treating epilepsy to each individual’s neurophysiology prior to clinical measures of complexity and least-square support vector machine. Informatics in
operation. Medicine Unlocked 2021. https://doi.org/10.1016/j.imu.2021.100536.
[19] Djoufack Nkengfack Laurent Chanel, Daniel Tchiotsop, Romain Atangana,
Saha Tchinda, Beaudelaire, Valérie Louis-Door, Wolf Didier. Jacobi polynomial
Ethical approval transforms-based entropy measures for focal and non-focal EEG signals
discrimination using kernel machines. Sci J Circ Syst Signal Process 2021;10(No.
This article does not contain any studies with human participants 2):25–37. https://doi.org/10.11648/j.cssp.20211002.11.
[20] Andrzejak RG, Lehnertz K, Mormann F, Rieke C, David P, Elger CE. Indications of
and/or animals performed by any of the authors. nonlinear deterministic and finite-dimensional structures in time series of brain
electrical activity: dependence on recording region and brain state. Phys Rev 2001;
64(No. 6). https://doi.org/10.1103/PhysRevE.64.061907.
[21] Pardey J, Robert S, Tarassenko L. A review of parametric modelling techniques for
Declaration of competing interest EEG analysis. Med Eng Phys January 1996;18(No. 1):2–11.
[22] Tchiotsop D, Wolf D, Louis-Dorr V, Husson R. ECG data compression using Jacobi
The authors declare that there is no conflict of interests regarding the polynomials. In: Proceedings of the 29th Annual International Conference of the
IEEE EMBS; 2007. p. 1863–7. Lyon France, August 23-26.
publication of this paper.
[23] Tchiotsop D, Ionita S. ECG data communication using Chebyshev polynomial
compression methods. Telecommunicatii Numere publicate. Asociaţia Generală a
Acknowledgements Inginerilor din România (AGIR), AN XVI; December 2010. p. 22–32. 2.
[24] Aurelle Tchagna Kouanou, Daniel Tchiotsop, René Tchinda, Christian Tchito
Tchapga, Nicole Kengnou Telem Adelaide, Romanic Kengne. A machine learning
The authors are very grateful to the anonymous referees for their Algorithm for biomedical images compression using orthogonal transforms. I.J.
valuable comments which helped to improve the quality of the paper. Image, Graphics and Signal Processing 2018;11:38–53. https://doi.org/10.5815/
ijigsp.2018.11.05.
[25] Kaiser JF. On a simple algorithm to calculate the energy of a signal. In: Proceedings
References of the International Conference on Acoustics, Speech and Signal Processing. IEEE;
1990. p. 381–4.
[26] Tzallas Alexandros T. Epileptic seizure detection in EEGs using time–frequency
[1] Porter RJ. Classification of epileptic seizures and epileptic syndromes. In:
analysis. IEEE Trans Inf Technol Biomed September 2009;13(No. 5).
Laidlaw J, Richens A, Chadwick D, editors. A Texbook of Epilepsy. New-York:
[27] Müller W, Jung A, Ahammer H. Advantages and problems of nonlinear methods
Churchill Livingstone; 1993. p. 1–22.
applied to analyze physiological time signals: human balance control as an
World health organization (WHO). at 02PM, https://www.who.int/news-room/fact-shee
example. Sci Rep 2017;7(No. 2464).
ts/detail/epilepsy. [Accessed 16 June 2021].
[28] Poomipat Boonyakitanont, Apiwat Lek-uthai, Krisnachai Chomtho, Jitkomut
[3] Bell GS, Sander JW. The Epidemiology of Epilepsy: The Size of the Problem’’. 2001.
Songsiri, ‘‘A review of feature extraction and performance evaluation in epileptic
p. 306–14. Seizure.
seizure detection using EEG’’, Biomed Signal Process Contr, Vol. 57, 2020. DOI:
[4] Subasi A, Gursoy MI. EEG signal classification using PCA, ICA, LDA and support
10.1016/j.bspc.2019.101702.
vector machines. Expert Syst Appl 2010;37:8659–66.
[29] Pearson Karl. On lines and planes of closest fit to systems of points in space.
[5] Nicolaou N, Georgiou J. Detection of epileptic electroencephalogram based on
Philosophical Magazine Series 1901;2(No. 11):559–72. 6.
permutation entropy and support vector machines. Expert Syst Appl 2012;39(No.
[30] Schölkopf B, Smola A, Müller KR. Nonlinear component analysis as a kernel
1):202–9.
eigenvalue problem. Neural Comput 1998:1299–319.
[6] Kumar Yatindra, Dewal ML, Anand RS. Epileptic seizures detection in EEG using
[31] Bishop Christopher M. Pattern Recognition and Machine Learning. Cambridge, U.
DWT-based ApEn and artificial neural network. SIViP 2014;(8):1323–34. https://
K.: Springer; 2006.
doi.org/10.1007/s11760-012-0362-9.
[32] Baudat G, Anouar F. Generalized discriminant analysis using a kernel approach.
[7] Zuhair Mohd, Thomas Sonia. Classification of patient by analyzing EEG signal
Neural Comput 2000;12:2385–404. https://doi.org/10.1162/
using DWT and least square support vector machine. Adv Sci Technol Eng Syst J
089976600300014980.
2017;2(No. 3):1280–9. 2415-669.
[33] Miller AS, Blott BH, Hames TK. Review of neural network applications in medical
[8] Li P, Karmakar C, Yearwood J, Venkatesh S, Palaniswami M, Liu C. Detection of
imaging and signal processing. Med Biol Eng Comput 1992;30:449–64.
epileptic seizure based on entropy analysis of short-term EEG. PloS One 2018;13
[34] Fausett L. Fundamentals of Neural Networks Architectures, Algorithms, and
(No. 3). https://doi.org/10.1371/journal.pone.0193691.
Applications. Englewood Cliffs, NJ: Prentice Hall; 1994.
[9] Paul Yash. Various epileptic seizure detection techniques using biomedical signals:
a review. Brain Informatics 2018;5(No. 6). https://doi.org/10.1186/s40708-018-
0084-z.

15
L.C. Djoufack Nkengfack et al. Informatics in Medicine Unlocked 26 (2021) 100721

[35] Orhan Umut, Hekim Mahmut, Ozer Mahmut. EEG signals classification using the K- domain features extraction and nonlinear analysis. Entropy 2017;19(No. 222).
means clustering and a multilayer perceptron neural network model. Expert Syst https://doi.org/10.3390/e19060222.
Appl 2011;38:13475–81. [46] Gupta Vipin, Ram Bilas Pachori. Epileptic seizure identification using entropy of
[36] Vladimir Vapnik. Statistical Learning Theory. Wiley and Sons; 1998. FBSE based EEG rhythms. Biomed Signal Process Contr 2019;53. https://doi.org/
[37] Suykens JA, Vandewalle J. Least squares support vector machine classifiers. Neural 10.1016/j.bspc.2019.101569.
Process Lett 1999;9:293–300. [47] Singh Gurwinder, Singh Birmohan, Kaur Manpreet. Grasshopper optimization
[38] Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating algorithm-based approach for the optimization of ensemble classifier and feature
characteristic (ROC) curve. Radiology 1982;143:29–36. selection to classify epileptic EEG signals. Med Biol Eng Comput 2019;57:1323–39.
[39] Kumar R, Indrayan A. Receiver operating characteristic (ROC) curve for medical https://doi.org/10.1007/s11517-019-01951-w.
researchers. Indian Pediatr 2011;48:277–87. [48] Khademul Islam Molla Md, Hassan Kazi Mahmudul, Islam Md Rabiul,
[40] Hajian-Tilaki K. Receiver operating characteristic (ROC) curve analysis for medical Tanaka Toshihisa. Graph eigen decomposition-based feature-selection Method for
diagnotic test evaluation. Caspian J. Intern. Med. 2013;4:627–35. epileptic seizure detection using electroencephalography. Sensors 2020;20(No.
[41] Suykens JA, Van Gestel T, De Brabanter J, De Moor B, Vandewalle J. Least squares 4639). https://doi.org/10.3390/s20164639.
support vector machines, vol. 4. World Scientific; 2002. [49] Li Mingyang, Sun Xiaoying, Chen Wanzhong, Jiang Yun, Zhang Tao. Classification
[42] Hasan Ocak. Automatic detection of epileptic seizures in EEG using discrete epileptic seizures in EEG using time-frequency image and block texture features.
wavelet transform and approximate entropy. Expert Syst Appl 2009;36(No. 2): IEEE ACCESS 2020;8:9770–81. https://doi.org/10.1109/ACCESS.2019.2960848.
2027–36. [50] Wu Jiang, Zhou Tengfei, Li Taiyong. Detecting epileptic seizures in EEG signals
[43] Song Yuedong, Lio Pietro. A new approach for epileptic seizure detection: sample with complementary ensemble empirical mode decomposition and extreme
entropy based feature extraction and extreme learning machine. J Biomed Sci Eng gradient boosting. Entropy 2020;22(No. 140). https://doi.org/10.3390/
2010;3:556–67. e22020140.
[44] Xiang Jie, Li Conggai, Li Haifang, Cao Rui, Wang Bin, Han Xiaohong, Chen Junjie. [51] Sameer Mustafa, Gupta Bharat. Detection of epileptical seizures based on alpha
The Detection of epileptic seizure signals Based on fuzzy entropy’’. J Neurosci band statistical features. Wireless Pers Commun 2020;115:909–25. https://doi.
Methods January 2015. https://doi.org/10.1016/j.jneumeth.2015.01.015. org/10.1007/s11277-020-07542-5.
[45] Wang Lina, Xue Weining, Yang Li, Luo Meilin, Huang Jie, Cui Weigang,
Huang Chao. Automatic epileptic seizure detection in EEG signals using multi-

16