EEG SIGNAL CLASSIFICATION USING LDA AND MLP CLASSIFIER

Health Informatics - An International Journal (HIIJ) Vol.9, No.1, February 2020
DOI:10.5121/hiij.2020.9102 14
EEG SIGNAL CLASSIFICATION USING LDA AND
MLP CLASSIFIER
Romain Atangana1,2,3
, Daniel Tchiotsop1
, Godpromesse Kenne1
, Laurent Chanel
DjoufackNkengfac k1,2
,
1Unité de Recherche d’Automatique et d’InformatiqueAppliquée (UR-AIA), IUT-FV of
Bandjoun, University of Dschang-Cameroon, P.O. Box 134-Bandjoun
2Unité 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
3Division of Continuing Education and Distance Learning, Higher Teacher Training
College, University of Ngaoundere-Cameroon, P.O. Box 652-Bertoua
ABSTRACT
Electroencephalography (EEG) is the recording of electrical activities along the scalp. EEG measures
voltage fluctuations resulting from ionic current flows within the neurons of the brain. Diagnostic
applications generally focus on the spectral content of EEG, which is the type of neural oscillations that
can be observed in EEG signal. EEG is most often used to diagnose epilepsy, which causes obvious
abnormalities in EEG readings. This powerful property confirms the rich potential for EEG analysis and
motivates the need for advanced signal processing techniques to aid clinicians in their interpretations.
This paper describes the application of Wavelet Transform (WT) for the processing of
Electroencephalogram (EEG) signals. Furthermore, the linear discriminant analysis (LDA) is applied for
feature selection and dimensionality reduction where the informative and discriminative two-dimension
features are used as a benchmark for classification purposes through a Multi-Layers Perceptron (MLP)
neural network. For five classification problems, the proposed model achieves a high sensitivity,
specificity and accuracy of 100%.Finally, the comparison of the results obtained with the proposed
methods and those obtained with previous literature methods shows the superiority of our approach for
EEG signals classification and automated diagnosis.
KEYWORDS
EEG, Discrete Wavelet Transform (DWT), Linear Discriminant Analysis (LDA), Multi-Layers Perceptron
(MLP)Classifier
1. INTRODUCTION
In clinical contexts, EEG refers to the recording of the brain’s spontaneous electrical activity
over a short period of time, usually 20 to 40 minutes as recorded from multiple electrodes
placed on the scalp. The brain’s electrical charge is maintained by billions of neurons. Neurons
are electrically charged (or “polarized’) by membrane transport proteins that pump ions across
their membranes. Neurons are constantly exchanging ions with the extracellular milieu for
example to maintain resting potentials and to propagate action potentials. Ions of similar charge
repel each other, and when many ions are pushed out of many neurons at the same time, they
can push their neighbors, who push their neighbors, and so on, in a wave. This process is known
as volume conduction. When the wave of ions reaches the electrodes on the scalp, they can push
electrons on the metal on the electrodes. Since metal conducts the push and pull of electrons
easily, the difference in push or pull voltages between any two electrodes can be measured by a
voltmeter. Recording these voltages over time gives us the EEG.

15
The electric potential generated by an individual neuron is far too small to be picked up by
EEG. EEG activity therefore always reflects the summation of the synchronous activity of
thousands or millions of neurons that have similar spatial orientation. If the cells do not have
similar spatial orientation, their ions do not line up and create waves to be detected. Pyramidal
neurons of the cortex are thought to produce the most EEG signal because they are well aligned
and fire together. Scalp EEG activity shows oscillations at a variety of frequencies. Several of
these oscillations have characteristics frequency ranges, spatial distributions and are associated
with different states of the brain functioning. These oscillations are referred to as sub-bands
based on the frequency ranges. These bands from low to high frequencies respectively are called
delta(𝛿) (range 0.5 - 4Hz), theta (θ) (range 4-8Hz), alpha (α) (range 8- 13Hz), beta (β) (range
13-32Hz), and gamma (γ) (range 32-45Hz) [1].
The visual distinction of seizure from common artifacts within an EEG measurement is based
on the shape and spikiness of the waveforms. A signal with seizure has a rhythmical and
prominent spike, whereas most of other artifacts are non-stationary and randomly shaped. But
considering the fact that the recorded EEG pattern is a special mapping of signals captured by
placement of electrodes onto different regions on the scalp, it is extremely difficult for human
beings to observe and understand the actual behavior of the brain by merely visual inspection.
Hence there is an ever-increasing demand of an easily accessible and fully automated epileptic
seizure detection system using EEG signals.
During this last decade, advancements in biomedical signal processing have led to the analysis
of EEG being increasingly numerical for automated diagnosis of brain diseases. Epilepsy is the
most prevalent neurological disorder in humans and is characterized by recurrent seizures [2].
Seizures are defined as sudden changes in the electrical functioning of the brain, resulting in
altered behaviors, such as losing consciousness, jerky movements, temporarily loss of breath
and memory loss. Most of epilepsy analysis methods developed in the 20th
century were based
on the concept of visual inspection of EEG signals by skilled electroencephalographers.
However, with the advancement in the field of signal processing and pattern recognition,
different automatic techniques of epileptic seizure have been developed in the last two decades
[3-4].
Although the occurrence of epileptic seizures seems unpredictable [3], exciting application of
seizure prediction technology in therapeutic treatment of epilepsy can be used to trigger
intervention in preventing seizures before they begin [5].
Recently, seizure prediction was extremely investigated through different models of automatic
EEG signals classification and seizure detection [6-24]. These models generally include three
steps and widely vary in their theoretical approaches of the problem, validation of results and
the amount of data analyzed. In early attempts to show a relationship between the raw EEG and
behavior, the first step is generally related to the decomposition of the raw data and the feature
extraction processes. The decomposition process is applied in order to explore the raw data in
frequency domain [19,23], in time-frequency domain [3, 10,15,16,18,22], and in multi-domain
[8,12,19,23]. After the decomposition process, statistic features and entropies features [6,
10,14,[20] were widely extracted. Since these extracted features are generally in high dimension
and are no more discriminative for different classes, the second step proposed a selection of an
optimal and low dimension features through a dimensionality reduction process. For this
purpose, principal component analysis (PCA), independent component analysis (ICA), LDA
and K-means clustering were applied [6]. Then, at the last step, these most prominent and
discriminative features are used as inputs of a kernel machine for the classification decision. For
this purpose, two mostly used kernel machines with different architectures were proposed, such
as artificial neural network (ANN) [3,7,9,11,13,21] and support vector machine (SVM)
[6,14,17,22]. Most of these studies focus on epilepsy by using a feature vector with at least three
dimensions as inputs of a complex kernel machine.

16
This paper applies WT to explore EEG signals in the time-frequency domain. EEG data are
projected into a set of wavelet coefficients that exhibit the correlation which exist between the
signals and their corresponding EEG rhythms. Furthermore, Areas under the frequency
spectrum of each wavelet coefficient are used as the features. These areas are calculated using
trapezoidal rule from numerical analysis. For feature extraction purposes we first transform each
wavelet coefficient into its time domain then we apply Fourier transform to the wavelet
coefficient to obtain the frequency spectrum [25, 26] LDA analysis is mainly used to classify
two class problems. The main task of LDA is to use hyper plane to separate or to discriminate
data. LDA is a predictable discriminant analysis method which enables us to predict the
membership of a sample of a class that has been predefined by a set of features. On a practical
point of view, this method finds a projection of original data on a more reduced space so that
features of different classes should be well separated.
The rest of this paper is organized as follows: the second section describes the EEG dataset
used. The EEG analysis using WT technique for rhythms separation is briefly presented. LDA
and MLP are also presented for the purpose of feature extraction and classification, respectively.
Simulation results are presented and discussed in section 3. Then, the last part summarizes this
work.
2. MATERIALS AND METHODS
2.1. Dataset description
In the last few years, most researchers used publicly available data described in [28] for their
research work in the field of epileptic seizure detection. We are also using the same benchmark
database in order to compare our results with the results of previous research works. The
database is prepared by taking inputs from different subjects and is divided into five Sets (A –
E) each containing 100 EEG samples recorded through single channel. The mental state of the
subjects in each data set at the time of data recording was as follows:
- A - B: Five healthy volunteers relaxed in an awake state with eyes open (A) and closed
(B);
- C- D: Activities measured during seizure free periods of EEG from five patients, all of
whom had achieved complete seizure control and were correctly diagnosed;
- E: Contains seizure activity.
The dataset was recorded using a 128-channel amplifier system and standardized 10-20
electrode placement scheme. After recording, the data were sampled and digitized at 173.61 Hz
in a time duration of 23.6 seconds using 12 bits resolution as shown in table 1. Therefore, the
frequency range with the characteristics mentioned above is from 0.05 – 86.80 Hz. For
extracting the feature from EEG samples, the frequency range of interest is 0-64 Hz.
The dataset used in this study was collected from ten patients with a sampling frequency of
173.61 Hz and consists of five sets as presented in Table 1.
Table 1: EEG dataset description [28].
Set
Number of
Segments
Record Mode Record Type
Number of Samples /
Duration (s)
SET A 100 Extracranial Normal (eyes opened) 4097 / 23.6

17
SET B 100 Extracranial Normal (eyes closed) 4097 / 23.6
SET C 100 Intracranial Pre-ictal 4097 / 23.6
SET D 100 Intracranial Pre-ictal 4097 / 23.6
SET E 100 Intracranial Epileptic 4097 / 23.6
Figure 1: Examples of EEG signals.
2.2. Wavelet transform (WT) analysis
EEG signals may contain non-stationary characteristics. Thus, efficient and instantaneous
separation of the EEG rhythms can be done based on the decomposition of the EEG signals into
wavelet coefficients.
Wavelet transform (WT) is a powerful spectral estimation technique for the time-frequency
analysis of a signal [29]. The continuous WT (CWT) of a signal ( )
x t is defined as:
*
,
1
( ) ( )
a b
t b
CWT x x t dt
a
a

−
 
 
 
=  (1)
Where ( )
t
 is called the mother wavelet function (MWT), while a and b ( ,
a b IR
 ) are
dilation and translation parameter, respectively.
The application of WT in engineering areas usually requires the discrete WT (DWT). DWT,
which is non-redundant and has less computational time than CWT. DWT is defined by using
the discrete values of the parameters a and b .
/2 /2
0 0 0
, ( ) ( ) ( )
m m
m n
DWT x x t t nb dt
a a

− −
−
=  (2)
Where m and n ( ,
m n Z
 ) are indicating frequency and time localization, respectively.
For a dyadic-orthonormal WT, 0
a and 0
b are set to 2 and 1, respectively. This choice provides
the basis for multi-resolution analysis. For the purpose, Mallat[30] developed a dyadic method

18
by which DWT separates a signal into low and high-frequency components via two finite
impulse response quadrature mirror filters, where g a high-pass filter is related to the MWT
and h is a low-pass filter related to the scaling function.
( ) ( 1) (1 )
n
g h h n
= − − (3)
( ) ( ) 2 (2 )
n
x h n x n
 
= −
 (4)
( ) ( ) 2 (2 )
n
x g n x n
 
= −
 (5)
The result transforms the original signal into two equal sub-bands /2
[0 ]
N
F
− and /2
[ ]
N N
F F
−
and the high-frequency component can be sent to the second stage to repeat the procedure as
presented by Figure 2. Finally, the signal is decomposed at the expected level. The MWT and
the wavelet decomposition level should be selected carefully.
Figure 2: Sub-bands separation using WT.
Table 2.Rhythms, corresponding levels and simulated frequencies for a four-level decomposition using
Db4 MWT.
EEG rhythm
Separated
signal
Frequency
range (Hz)
Simulated frequency (Hz)
Set A EEG Set E EEG
Delta (δ) A4 0-4 0.21 3.60
Theta (θ) D4 4-8 5.76 5.97
Alpha (α) D3 8-16 11.53 14.11
Beta (β) D2 16-32 26.10 26.19
Gamma (γ) D1 32-64 50.00 64.33
Figure 3: Wavelet coefficients for the EEG rhythms separation of (a) a normal EEG Z001 and (b) an ictal
EEG S001.

19
After rhythms separation with the WT, the frequency distribution of the EEG signal is analyzed
by calculation of areas under the frequency spectrum of each wavelet coefficient through the
trapezoidal rule. Trapezoidal rule aims to approximate integral ∫ 𝑓(𝑥) 𝑑𝑥
𝑏
𝑎
by the sum of the
“areas” of the trapezoids as shown in Figure 4.
Figure 4: Trapezoidal approximation to ∫ 𝑓(𝑥) 𝑑𝑥
𝑏
𝑎
∫ 𝑓(𝑥)𝑑𝑥
𝑏
𝑎
=
1
2
(𝑦0 + 𝑦1)ℎ +
1
2
(𝑦1 + 𝑦2)ℎ + ⋯ +
1
2
(𝑦𝑛−1 + 𝑦𝑛)ℎ (5)
In which 𝑦𝑗 = 𝑓(𝑥𝑗)𝑓𝑜𝑟𝑡ℎ𝑒𝑐ℎ𝑜𝑜𝑠𝑒𝑛𝑝𝑎𝑟𝑡𝑖𝑡𝑖𝑜𝑛𝑎 = 𝑥0 < 𝑥1 < 𝑥2 < ⋯ < 𝑥𝑛−1 < 𝑥𝑛 = 𝑏.
And h is the spacing between successive points. If we set, ℎ =
(𝑏−𝑎)
𝑛
then the trapezoidal rule
can be written as:
∫ 𝑓(𝑥)𝑑𝑥
𝑏
𝑎
=
(𝑏−𝑎)
2𝑛
(𝑦0 + 2𝑦1 + 2𝑦2 + ⋯ + 2𝑦𝑛−1 + 𝑦𝑛) (6)
2.3. Linear discriminant analysis (LDA) feature Selection
Once that we have our feature vector, in order to increase the performance of the classifier it is
advisable to reduce its dimensionality by keeping the representative information in the feature
vector and eliminating the redundant one from it. The resulting vector is called reduced feature
vector. There are two main strategies of dimensionality reduction.
- Feature projection: This strategy tries to determine the best combination of the original
features to form a new feature set, generally smaller than the original one [31]. Principal
Component Analysis (PCA) is using this type of strategy, it is an orthogonal linear
transformation that can be used for dimensionality reduction in a signal database by
retaining those characteristics of the dataset that contribute most to its variance in order
to do this; lower order principal components are kept and higher order ones are
dismissed [32].This could be applied to EEG signals.
- Feature Selection: This strategy chooses the best subset of the original feature vector
according to some criteria for judging whether one subset is better than another. The
ideal criterion for classification should be the minimization of the probability of
misclassification, but generally simpler criteria based on class separability are chosen
[31] Linear Discriminant Analysis (LDA) is used in that type of strategy.

20
LDA analyzes signals to extract the most prominent features. The main aim of LDA feature
extraction is to obtain further information from the signal by transforming the feature vector
formed by the areas into fewer ones as a feature vector. The extracted features are more
informative and discriminative with reducednoise and redundancy compared to the original
data.
LDA is a supervised algorithm since it computes the linear discriminant features by maximizing
distance between classes and minimizing distance within classes [6]. The LDA projection can
be written as a single matrix equation described by:
*
Y X W
= (7)
where Y and X are the original data and features, respectively. And W is the dxk projection
matrix formed of the first k-eigenvectors of 1
w b
S S
−
, where w
S and b
S are the within-class and
between-class scatter matrices, respectively.
( ) ( )
1 1
1 l
N
L
T
l l
w i l i l
l i
S x x
N
 
= =
= − −
 (8)
( ) ( )
1
1 L
T
b l l l
l
S N
N
   
=
= − −
 (9)
l
i
x
, l
N
and l
 are an i-row vector, the number of i-vectors and the mean vector corresponding
to class label l, respectively. L, N and
 are the number of classes, the total number of i-
vectors and the global mean of all classes, respectively.
2.4. Multi-Layer Perceptron (MLP) Classifier
2.4.1. Architecture
Multi-Layer perceptron (MLP) represents a generalization of the single layer perceptron. A
single layer perceptron forms half -plane decision regions, while multi – layer perceptron can
form arbitrarily complex decision regions and can separatevarious input patterns [27]. Figure 5
shows a typical multi-layer perceptron neural network structure.

21
Figure 5: Multi-Layer Perceptron classifier.
In the context of biomedical signal processing, especially with application to EEG signals, the
classification of data in feature spaces is often required. The objective of classification is to
draw a boundary between two or more classes and to label them based on their measured
feature, the task being to assign an input pattern represented by a feature vector to one or many
pre-specified classes.
The MLP is a popular learningmachine for pattern classification[2,11, 28, 29]. In this paper, the
goal is to classify different sets of EEG signals using MLP. MLP consists of the following
layers.
Input layer: A layer of neurons that receives information from external sources and passes this
information to the network for processing.
Hidden layer: A layer of neurons that receives information from the input layer and processes
them in a hidden way. It has no direct connection withother layers within the system.
Output layer: A layer of neurons that receives processed information and sends output signal out
of the system.
Bias: Acts on the neuron like an offset. The function of the bias is to provide a threshold for the
activation of neurons. The bias input is connected to each of the hidden and output neurons in a
network.
The ability to learn is a fundamental property of intelligence. Although a precise definition of
learning is difficult to formulate, a learning process in the MLP can be viewed as the problem of
updating connection weights so that a network can efficiently perform a specific task. The
network usually must learn the connection weights from available training patterns.
Performance is improved over time by iteratively updating the weights in the network. ANNs
ability to automatically learn from examples makes them attractive and exciting. Instead of
following a set of rules specified by human experts, ANNs appear to learn underlying rules from
the given collection of representative examples. This is one of the major advantages of neural
networks over traditional expert systems [30].
2.4.2. Learning process
There are three main learning paradigms: Supervised, Unsupervised, and hybrid. In Supervised
learning, or learning with a “teacher”, the network is provided with a correct answer (output) for
every input pattern. Weights are determined to allow the network to provide answers as close as
possible to the known correct answers. Examples of the supervised learning algorithms are the
least mean square (LMS) algorithm and its generalization, known as the back-propagation
algorithm[31-34]. In this work, the sigmoid function is used as activation function. The weights
and bias are determined using the back-propagation algorithm, which is based on searching a
minimum mean square error between the desired and actual outputs using gradient descent [29].
In contrast, unsupervised learning or learning without a teacher does not require a correct
answer associated with each input pattern in the training data set. It explores the underlying
structure in the data or correlation between patterns in the data and organizes patterns into
categories from correlations. Kohonen network and K-means algorithm are examples of
unsupervised learning.
Hybrid learning combines supervised and unsupervised learning. Part of the weights is usually
determined through supervised learning, whileothers are obtained through unsupervised
learning.

22
2.4.3. Back Propagation Algorithm
Backward of the output error
Learning by gradient descent
- Output layer:Correction guided by the error between the desired and the actual output;
- Hidden layers: Correction about the sensitivity (influence of the neuron on the error in
the output layer).
Figure 6:Mathematical Model of ANN.
Output values of the neurons
Value𝑦𝑗
𝑡
of neuron j for the input data𝑋𝑡
𝑦𝑗
𝑡
= 𝑓(𝑎𝑗
𝑡
) = 𝑓(∑ 𝜔𝑗,𝑖
𝑅
𝑖=1 . 𝑦𝑖
𝑡
+ 𝜔𝑗,0) (10)
f: activation function of the neuron;
𝑎𝑗
𝑡
= ∑ 𝜔𝑗,𝑖
𝑅
𝑖=1 . 𝑦𝑖
𝑡
+ 𝜔𝑗,0is the cumulative sum of input neurons;
𝜔𝑗,𝑖: weights of the link connecting neuron j to neuron i of the previous layer;
𝜔𝑗,0: bias of neuron j;
𝑦𝑖
𝑡
: output of neuron iof the previous layer for the input data 𝑋𝑡
;
𝑅: numberof neurons on the previous layer.
Output error to the output layer
A set of data χ = {𝑥𝑡
, 𝑟𝑡}𝑡=1
𝑁
with 𝑟𝑡
= [𝑟1
𝑡
𝑟2
𝑡
… . . 𝑟𝐾
𝑡 ]𝑇
=
[
𝑟1
𝑡
𝑟2
𝑡
.
.
𝑟𝐾
𝑡
]
Where𝑟𝑗
𝑡
= 1 𝑖𝑓𝑥𝑡
𝜖𝐶𝑗, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒𝑟𝑗
𝑡
= 0
Error observed for the data 𝑥𝑡
on neuron j on the output layer
𝑒𝑗
𝑡
= 𝑟𝑗
𝑡
− 𝑦𝑗
𝑡
(11)

23
Quadratic error observed for the data 𝑥𝑡
on all the K neurons of the output layer (one neuron per
class).
𝐸𝑡
=
1
2
∑ (𝑒𝑗
𝑡
)
2
𝐾
𝑗=1 (12)
Quadratic mean error observed for the data set χ.
𝐸 =
1
𝑁
∑ 𝐸𝑡
𝑁
𝑡=1 (13)
Correction of the error for the output layer
Correction of weights through descent gradient of the quadratic mean error
∆𝜔𝑗,𝑖 = −𝜂
𝜕𝐸
𝜕𝜔𝑗,𝑖
= −𝜂
𝜕(
1
𝑁
∑ 𝐸𝑡
𝑁
𝑡=1 )
𝜕𝜔𝑗,𝑖
(14)
Δ𝜔𝑗,𝑖 = −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝜔𝑗,𝑖
𝑁
𝑡=1
The error of neuron j depends on neurons of the previous layer
Expanding using the chain derivative rule.(
𝜕𝑓
𝜕𝑥
=
𝜕𝑓
𝜕𝑦
𝜕𝑦
𝜕𝑥
)
𝜕𝐸𝑡
𝜕𝜔𝑖,𝑗
=
𝜕𝐸𝑡
𝜕𝑒𝑗
𝑡
𝜕𝑒𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,𝑖
(15)
𝜕𝐸𝑡
𝜕𝜔𝑗,0
=
𝜕𝐸𝑡
𝜕𝑒𝑗
𝑡
𝜕𝑒𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝜕𝑎𝑗
𝑡
ð𝜔𝑗,0
(16)
Calculation of the partialderivatives
Expanding with the sigmoid activation function𝑦𝑗
𝑡
=
1
1+exp (−𝑎𝑗
𝑡
)
(17)
ð𝐸𝑡
𝜕𝑒𝑗
𝑡 =
𝜕
𝜕𝑒𝑗
𝑡 [
1
2
∑ (𝑒𝑗
𝑡
)
2
𝐾
𝑗=1 ] = 𝑒𝑗
𝑡
(18)
𝜕𝑒𝑗
𝑡
𝜕𝑦𝑗
𝑡 =
𝜕[𝑟𝑗
𝑡
−𝑦𝑗
𝑡
]
𝜕𝑦𝑗
𝑡 = −1 (19)
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡 =
𝜕[
1
1+𝑒𝑥𝑝(− 𝑎𝑗
𝑡)
]
𝜕𝑎𝑗
𝑡 =
𝑒𝑥𝑝(−𝑎𝑗
𝑡
)
[1+𝑒𝑥𝑝(−𝑎𝑗
𝑡
)]
2 = [
1
1+𝑒𝑥𝑝(−𝑎𝑗
𝑡
)
] [1 − {
1
𝑡
)
}](20)
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡 =
1
𝑡
)
∗
𝑒
−𝑎𝑗
𝑡
+1−1
1+𝑒𝑥𝑝(−𝑎𝑗
𝑡
)
= 𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
) (20)
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡 = 𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
) (20)
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,𝑖
=
𝜕
𝜕𝜔𝑗,𝑖
∑ 𝜔𝑗,𝑖𝑦𝑖
𝑡
𝑅
𝑖=1 + 𝜔𝑗.0 = 𝑦𝑖
𝑡
(21)
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,0
= 1 (22)
Learning for the output layer
Adjusting the weights of the output layer
∆𝜔𝑗,𝑖 = −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝜔𝑗,𝑖
𝑁
𝑡=1 = −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝑒𝑗
𝑡
𝜕𝑒𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,𝑖
𝑁
𝑡=1 (23)

24
∆𝜔𝑗,𝑖 = +
𝜂
𝑁
∑ 𝑒𝑗
𝑡
𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
)𝑦𝑖
𝑡
𝑁
𝑡=1 (24)
Δ𝜔𝑗,0 = −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝜔𝑗,0
𝑁
𝑡=1 = −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝑒𝑗
𝑡
𝜕𝑒𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,0
𝑁
𝑡=1 (25)
∆𝜔𝑗,0 =
𝜂
𝑁
∑ 𝑒𝑗
𝑡
𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
)
𝑁
𝑡=1 (26)
Delta Rule
Let a delta 𝛿𝑗
𝑡
, which corresponds to the local gradient of neuron j for the input data 𝑋𝑡
.
𝛿𝑗
𝑡
= 𝑒𝑗
𝑡
𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
) (27)
ThusΔ𝜔𝑗,𝑖 =
𝜂
𝑁
∑ 𝛿𝑗
𝑡
𝑦𝑖
𝑡
𝑁
𝑡=1 (28)
And Δ𝜔𝑗,0 =
𝜂
𝑁
∑ 𝛿𝑗
𝑡
𝑁
𝑡=1 (29)
Correction of the error for the hidden layers
Gradient of the error for the hidden layers.
𝜕𝐸𝑡
𝜕𝜔𝑗,𝑖
=
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,𝑖
(30)
Only
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡changes, while
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡 𝑎𝑛𝑑
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,𝑖
are the same as on the output layer.
The error for a neuron of the hidden layer depends on the error of the K neurons of the next
layer. (Back-propagation of the error)
𝐸𝑡
=
1
2
∑ (𝑒𝑘
𝑡)2
𝑁
𝑘=1 (31)
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡 =
𝜕
𝜕𝑦𝑗
𝑡 [
1
2
∑ (𝑒𝑘
𝑡)2
𝑘 ] = ∑ 𝑒𝑘
𝑡 𝜕𝑒𝑘
𝑡
𝜕𝑦𝑗
𝑡
𝑘 (32)
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡 = ∑ 𝑒𝑘
𝑡 𝜕𝑒𝑘
𝑡
𝜕𝑎𝑘
𝑡
𝜕𝑎𝑘
𝑡
𝜕𝑦𝑗
𝑡
𝑘 = ∑ 𝑒𝑘
𝑡 𝜕[𝑟𝑘
𝑡
−𝑦𝑘
𝑡
]
𝜕𝑎𝑘
𝑡
𝜕[∑ 𝜔𝑘,𝑙𝑦𝑘
𝑡
+𝜔𝑘,0
𝑙 ]
𝜕𝑦𝑗
𝑡
𝑘 (33)
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡 = ∑ 𝑒𝑘
𝑡
𝑘 [−𝑦𝑘
𝑡(1 − 𝑦𝑘
𝑡)]𝜔𝑘,𝑗 (33)
𝛿𝑘
𝑡
= 𝑒𝑘
𝑡[𝑦𝑘
𝑡(1 − 𝑦𝑘
𝑡)] (34)
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡 = − ∑ 𝛿𝑘
𝑡
𝑘 𝜔𝑘,𝑗 (35)
Correction of the corresponding error
𝜕𝐸𝑡
𝜕𝜔𝑗,𝑖
=
𝜕𝐸𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑦𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝑎𝑗
𝑡
𝜕𝜔𝑗,𝑖
= − [∑ 𝛿𝑘
𝑡
𝜔𝑘,𝑗
𝑘 ]𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
)𝑦𝑖
𝑡
(36)
𝛿𝑗
𝑡
= 𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
)∑ 𝛿𝑘
𝑡
𝑘 𝜔𝑘,𝑗 (37)
∆𝜔𝑗,𝑖 = −𝜂
𝜕𝐸
𝜕𝜔𝑗,𝑖
= −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝜔𝑗,𝑖
𝑁
𝑡=1 =
𝜂
𝑁
∑ 𝛿𝑗
𝑡
𝑦𝑖
𝑡
𝑁
𝑡=1 (38)
Δ𝜔𝑗,0 = − 𝜂
𝜕𝐸
𝜕𝜔𝑗,0
= −
𝜂
𝑁
∑
𝜕𝐸𝑡
𝜕𝜔𝑗,0
𝑁
𝑡=1 =
𝜂
𝑁
∑ 𝛿𝑗
𝑡
𝑁
𝑡=1 (39)

25
Figure 7: Back-propagation of error.
The back-propagation algorithm can be summarized by the following steps:
Step1:Normalize the training data𝑋𝑖
𝑡
𝜀[−1,1]and the desired outputs𝑟𝑖
𝑡
̃ 𝜀 {0,05 ; 0,95};
Step 2: Initialize the weights and bias to small random values, 𝜔𝑗,𝑖 𝜀 𝒰(−0,5 ,0,5) ;
Step3: Select the structure of the network (such as the number of hidden layers and number of
neurons for each layer);
Step4: Choose activation functions for the neurons. These activation functions can be uniform,
or they can be different for different layers;
Step5: Select the training pair from the training set. Apply the input vector to the network input
and propagate the signal forward through the network;
Step6: Calculate the output of the network based on the initial weights and input set;
Step7: Calculate the error between network output and desired output;
𝑒𝑗
𝑡
= 𝑟̃𝑗
𝑡
− 𝑦𝑗
𝑡
; 𝑗 = 1 … … . . , 𝐾 ; 𝑡 = 1, … … … . , 𝑁 (40)
Step8: Propagate error backward and adjust the weights in such a way that minimizes the error.
Start from the output layer and go backward to the input layer;
𝜔𝑗,𝑖 = 𝜔𝑗,𝑖 + ∆𝜔𝑗,𝑖 = 𝜔𝑗,𝑖 +
𝜂
𝑁
∑ 𝛿𝑗
𝑡
𝑡 𝑦𝑖
𝑡
(41)
𝜔𝑗,0 = 𝜔𝑗,0 + Δ𝜔𝑗,0 = 𝜔𝑗,0 +
𝜂
𝑁
∑ 𝛿𝑗
𝑡
𝑡 (42)
Where the local gradient or Delta is defined as:
𝛿𝑗
𝑡
= {
𝑒𝑗
𝑡
𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
) , 𝑖𝑓 𝑗 𝑏𝑒𝑙𝑜𝑛𝑔𝑠 𝑡𝑜 𝑡ℎ𝑒 𝑜𝑢𝑡𝑝𝑢𝑡 𝑙𝑎𝑦𝑒𝑟
𝑦𝑗
𝑡
(1 − 𝑦𝑗
𝑡
) ∑ 𝛿𝑘
𝑡
𝜔𝑘,𝑗
𝑘 , 𝑖𝑓 𝑗 𝑏𝑒𝑙𝑜𝑛𝑔𝑠 𝑡𝑜 𝑡ℎ𝑒 ℎ𝑖𝑑𝑑𝑒𝑛 𝑙𝑎𝑦𝑒𝑟
(43)
Step9: Repeat steps 5-8 for each vector in the training set until the error for the set is lower than
the required minimum error.
After enough repetitions of these steps the error between the actual output and the target output
should be reduced to an acceptable value, and the network is said to be trained. At this point the
network can be used in the recall or generalization phase where the weights are not changed.
2.5. Classification Problems and Performances Evaluation
2.5.1. Terminology and Standard Measures
The American Heart Association (AHA) recommends the use of four measures to evaluate
procedures for diagnosing CAD [35]. Since these measures are useful in other areas of diagnosis
as well, we will be using them in evaluating epilepsy diagnosis system.
Sensitivity …. 𝑇𝑃𝐹 =
𝑇𝑃∗100
𝑇𝑃+𝐹𝑁
(44)

26
Specificity….𝑇𝑁𝐹 =
𝑇𝑁∗100
𝑇𝑁+𝐹𝑃
(45)
𝑃𝐴 = 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 ∗ 𝑃(𝐷) + 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 ∗ [1 − 𝑃(𝐷)] (46)
𝑃𝑉 =
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦∗𝑃(𝐷)
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦∗𝑃(𝐷)+(100−𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦)∗[1−𝑃(𝐷)]
(47)
𝐴𝐶𝐶 =
𝑇𝑃+𝑇𝑁
𝑇𝑃+𝐹𝑁+𝑇𝑁+𝐹𝑃
∗ 100 (48)
Where TP stands for true positive, FN stands for false negative, TN stands for true negative, and
FP stands for false positive.
Sensitivity, or true-positive fraction (TPF), is the probability of a patient who is suffering from
an epilepsy to be diagnosed as such.
Specificity, or true negative fraction (TNF), is the probability that a healthy individual is
diagnosed as such by an epilepsy diagnosis mechanism.
PA is the predictive accuracy or overall percentage of correct diagnosis.
PV is the predictive value of a positive test or the percentage of those who have the disease and
have tested positive for it.
P(D) is the a priori probability of a patient who is referred to the diagnosis procedure
having epilepsy.
Accuracy (ACC): sum of correctly classified patterns (TP TN
+ )/sum of correctly and non-
correctly classified patterns (TP FN TN FP
+ + + ).
In addition to TPF and TNF, we define two other related values.
False positive fraction (FPF) is the probability of a healthy patient being incorrectly
diagnosed as suffering from epilepsy.
False negative fraction (FNF) is the probability that a patient who is suffering from epilepsy will
be incorrectly diagnosed as healthy. In this way, the following relations can be established:
𝐹𝑃𝐹 = 1 − 𝑇𝑁𝐹 (49)
𝐹𝑁𝐹 = 1 − 𝑇𝑃𝐹 (50)
In this paper, five classification problems that are closely related to the clinical applications are
designed from the entire database.
Table 3. Classes and corresponding number of EEG for five classification problems.
Classification
problem
Classes
Number of EEG
segments
A – E
Normal (A)
Ictal (E)
100
100
B – E
Normal (B)
Ictal (E)
100
100
AB - E
Normal (A, B)
Ictal (E)
200
100
AB - CD
Normal (A, B)
Pre-ictal (C, D)
200
200
CD-E
Pre-ictal (C, D)
Ictal (E)
200
100

27
The sensitivity, specificity, and accuracy parameters are used as performance metrics of the
MLP for each classification problem [4].
3. RESULTS AND DISCUSSION
This section presents and discusses the experimental results of the proposed model in order to
determine its capability and consistency. All experiments have been carried out in JAVA. In this
study, all the database is used to perform different classification problems. Firstly, EEG rhythms
are separated using WT and the frequency distribution of the EEG signal is analyzed in terms of
the areas under the frequency spectrum of each wavelet coefficient. According to their
supervision, it is observed that areas analysis of the database is representative for the various
classes of signals but still presents redundancy. Therefore, the LDA is applied to select and
extract suitable, easily viewed and interpreted two-dimension features. Figure 5 presents how
these extracted 2-dimension feature vectors represent the amount of changes in frequency
distribution and carry discriminative and representative information about their classes for
different classification problems. Then, these feature vectors can be used as benchmark of
classification research through the proposed MLPN classifier. Compared to others, these feature
vectorsare extremely discriminative since more of the previous proposed models use complex
kernel machines after feature extraction. For different classification problems, the first 40% of
each SET are used to determine the useful linear discriminants for the feature extraction
process, the remaining 60 % of each set is used for the classification process.

28
Figure 5: Scatter plot of LDA features for different classification problems.
Therefore, the performances of the MLPNN for the five classification problems are evaluated
and summarized in Table 4 in terms of sensitivity, specificity and accuracy. In the first two
classifications 𝜂 = 0.8, and 𝜂 = 0.6 for the rest. The first 70% of each SET are used as the
training set and the last 30% is divided into two parts; 15 % as the validation set and 15 % for
the testing set. In order to achieve a good generalization or convergence, the input features and
the output target are normalized.
A-E (ɳ= 0.8)
B-E (ɳ= 0.8)
AB-CD (ɳ= 0.8)

29
CD-E (ɳ= 0.6)
Figure 6: Convergence of training and validation error of the MLPNN with 1, 12, 12 and 2 neurons at the
input layer, first hidden layer, second hidden layer and output layer, respectively.
Table 4: Performances of the MLPNN for five classification problems.
Classification
Problem
MLPNN Performances (%)
Sensitivity Specificity Accuracy
A - E 100 100 100
B - E 100 100 100
AB - E 100 100 100
CD - E 100 100 100
AB - CD 100 100 100
For A-E, B-E, AB-E, CD-E and AB-CD classification problems, the proposed model achieved a
maximum accuracy of 100%. Then, for this clinical diagnosis, the proposed model can be used,
and good decision can be taken out.
Compared to other works, in general, for different experiments; the model proposed in this
paper achieves a high accuracy. The MLPNN architecture used in this work has a lower
computational cost and is faster to compute. Then, this WT-LDA-MLPNN model can be
used for EEG signals classification problems since highly discriminative features are extracted
and high accuracy is obtained.
4. CONCLUSION
In this paper, WT has been applied for time-frequency decomposition and their rhythm
distributions have been analyzed using areas under the frequency spectrum of each wavelet
coefficient. Furthermore, LDA has been used to extract discriminative 2-dimension feature
vectors as input of a MLPNN. In terms of their sensitivity, specificity and accuracy, the
performances for five classification problems were derived. The result of EEG signal
classification through MLPNN has demonstrated that LDA feature extraction can improve the
performance of classifiers with respect to reducing the computational cost. According to this
result, the proposed model can serve as a promising alternative for intelligent diagnosis system.
REFERENCES
[1] SaeidSanei and J.A. Chambers., [2007] “EEG signal Processing”. John Wiley & Sons.
[2] Blume W., Lüders H., Mizrahi E., Tassinari C., van Emde Boas W., Engel J. [2001], “Glossary of
Descriptive Terminology for Ictal Semiology: Report of the ILAE Task Force on Classification and
Terminology”, Epilepsia, Vol. 42, No. 9, pp. 1212-1218
[3] Subasi Abdulhamit [2007], ‘‘EEG Signal Classification using Wavelet Feature Extraction and a
Mixture of Expert Model’’, Elsevier Transactions on Expert System with Applications, Vol. 32, No.
5, pp. 1084-1093.

30
[4] K. Fu, J.Qu, Y. Chai, and Y. Dong [2014] “Classification of seizure based on the time – frequency
image of EEG signals using HHT and SVM”. Biomedical Signal Processing and Control vol -13,
pp: 15-22.
[5] Adeli H., Zhou Z. and Dadmehr N. [2003] “Analysis of EEG Records in an Epileptic Patient Using
Wavelet Transform”, Journal of Neuroscience Methods, Vol. 123, pp. 69-87.
[6] Subasi A., Gursoy M. I. [2010], ‘‘EEG Signal Classification Using PCA, ICA, LDA and Support
Vector Machines’’, Expert Systems with Applications, Vol. 37, pp. 8659-8666.
[7] Tzallas A. T., Tsipouras M. G., Fotiadis D. I. [2007], “Automatic Seizure Detection Based on Time-
Frequency Analysis and Artificial Neural Networks”,Comput. Int. Neurosci., pp. 1-13. DOI:
10.1155/2007/80510.
[8] Pachori R. B., Bajaj V. [2008], ‘‘Discrimination between Ictal and Seizure-Free EEG Signals Using
Empirical Mode Decomposition’’, Res. Lett. Signal Process. DOI: 10.1155/2008/293056
[9] Sivasankari N. and Thanushkodi K. [2009],“Automated Epileptic Seizure Detection in EEG Signals
Using FastICA and Neural Network”, Int. J. Advance. Soft Comput. Appl., Vol. 1, No. 2,
November. ISSN: 2074-8523
[10] Ocak H. [2009], ‘‘Automatic Detection of Epileptic Seizures in EEG Using Discrete Wavelet
Transform and Approximate Entropy’’, Expert Systems with Applications, Vol. 36, No. 2, pp. 2027-
2036.
[11] Ling Guo, Daniel Rivero, Julián Dorado, Juan R. Rabunal, Alejandro Pazos [2010], “Automatic
Epileptic Seizure Detection in EEG Based on Line Length Feature and Artificial Neural Networks”,
Journal of Neuroscience Methods, Vol. 191, pp. 101-109. DOI: 10.1016/j.jneumeth.2010.05.020.
[12] Pachori R. B. and Bajaj V. [2011], ‘‘Analysis of Normal and Epileptic Seizure EEG Signals Using
Empirical Mode Decomposition’’, Comput. Methods Progr. Biomed., Vol. 104, pp. 373-381.
[13] OrhanUmut, HekimMahmut, Ozer Mahmut [2011], “EEG Signals Classification Using the K-
Means Clustering and a Multilayer Perceptron Neural Network Model”, Expert Systems with
Applications, Vol. 38, pp. 13475-13481..
[14] Nicolaou N., Georgiou J. [2012], ‘‘Detection of Epileptic Electroencephalogram Based on
Permutation Entropy and Support Vector Machines’’, Expert Systems with Applications, Vol. 39,
No. 1, pp. 202-209.
[15] Saching Garg, Rakesh Narvey [2013], ‘‘Denoising and Feature Extraction of EEG Signal Using
Wavelet Transform’’, International Journal of Engineering Science and Technology (IJEST), Vol. 5,
No. 6, pp. 1249-1253, June. ISSN: 0975-5462.
[16] Rosenblatt Mariel et al. [2014], ‘‘A Quantitative Analysis of an EEG Epileptic Record Based on
Multiresolution Wavelet Coefficients’’, Entropy, Vol. 16, pp. 5976-6005, 2014.
[17] Kai Fu, Qu J., Chai Y., Dong Y. [2014], ‘‘Classification of Seizure Based on the Time-Frequency
Image of EEG Signals Using HHT and SVM’’, Biomedical Signal Processing and Control, Vol. 13,
pp. 15-22.
[18] Ashwini Nakate, P. D. Bahirgonde [2015], ‘‘Feature Extraction of EEG Signal UsingWavelet
Transform’’, International Journal of Computer Applications,Vol. 124, No. 2, August.
[19] Pushpendra Singh, S. D. Joshi,·R. K. Patney, Kaushik Saha [2016], “Fourier-Based Feature
Extraction for Classification of EEG Signals Using EEG Rhythms”,Circuits Syst Signal Process,
Vol. 35, pp. 3700-3715. DOI: 10.1007/s00034-015-0225-z.
[20] Ramakrishnan J, Bommanna Raja Kanagaraj [2017], “Analysis of Non-seizure and Seizure Activity
Using Intracranial EEG Signals and Empirical Mode Decomposition Based Approximate
Entropy”,Biomedical Research, Vol. Special Issue. ISSN: 0970-938X.
[21] Sandeep Kumar Satapathy, SatchidanandaDehuri, Alok Kumar Jagadev [2017], “EEG Signal
Classification Using PSO Trained RBF Neural Network for Epilepsy Identification”, Informatics in
Medicine Unlocked, Vol. 6, pp.1-11.
[22] Mohd Zuhair, Sonia Thomas [2017], ‘‘Classification of Patient by Analyzing EEG Signal Using
DWT and Least Square Support Vector Machine’’, Advances in Science, Technology and
Engineering Systems Journal, Vol. 2, No. 3, pp. 1280-1289, 2017. ISSN: 2415-669.
[23] Subasi Abdulhamit, Kevric Jasmin, Canbaz M. Abdullah [2017], “Epileptic Seizure Detection
Using Hybrid Machine Learning Methods”, Neural Comput and Applic. DOI: 10.1007/s00521-017-
3003-y.
[24] Wang Lina, XueWeining, Li Yang, Luo Meilin, Huang Jie, Cui Weigang and Huang Chao [2017],
“Automatic Epileptic Seizure Detection in EEG Signals Using Multi-Domain Feature Extraction
and Nonlinear Analysis”, Entropy, Vol. 19, No. 222. DOI: 10.3390/e19060222
[25] Jérôme Landre et al., [2005] “Analyse multiresolution pour la recherché et l’indexation d’image par
le contenu dans les bases de donnees image”. Application a la base d’images paléontologiques
Trans’ Tyfipal; Interface homme-machine. Thèse de Doctorat de l’Université de Bourgogne.

31
[26] PulomaPramanick et al., [2013] ‘Classification of Electroencephalogram (EEG) signal based on
Fourier Transform and neural network’. Mémoire de recherche au Génie Electrique de l’Institut
National de Technologie de Rourkela.
[27] Nicolaou N., Georgiou J. [2012], ‘‘Detection of Epileptic Electroencephalogram Based on
Permutation Entropy and Support Vector Machines’’, Expert Systems with Applications, Vol. 39,
No. 1, pp. 202-209.
[28] Andrzejak R. G., Lehnertz K., Mormann F., Rieke C., David P. and Elger C. E. [2001],
‘‘Indications of Nonlinear Deterministic and Finite-Dimensional Structures in Time Series of Brain
Electrical Activity: Dependence on Recording Region and Brain State’’, Physical Review E., Vol.
64, No. 6. DOI: 061907.
[29] Daubechies I. [1992], ‘‘Ten Lectures on Wavelets’’, Industrial and Applied Mathematics,
Philadelphia
[30] Mallat S. [1989], ‘‘A Theory for Multi-resolution Signal decomposition: The wavelet
Representation’’, IEEE Transaction on Analy. Mach. Intell., Vol. 11, No. 7, pp. 674-693.
[31] M. Zecca, S. Micera, MC. Carrozza, and P. Dario [2002] ”Control of multifunctional prosthetic
hands by processing the electromyography signal’. Critical Reviews in Biomedical Engineering,
citeeer, 30(4-60; 459.
[32] A. Bashashati, M. Fatourechi, R. K Ward, and G. E Birch [2007]. “A Survey of signal processing
algorithms in brain- computer interface based on electrical brain signals”. Journal of Neural
engineering 4; R32.
[33] Ali Zilouchian and Mohammad Jamshidi : Intelligent control systems using soft computing
methodologies.p. cm.Includes bibliographical references and index.ISBN 0-8493-1875-01.
Intelligent control systems—Data processing. 2. Soft computing. TJ217.5 .I5435 2001 629.89′
0285′63—dc21
[34] Miller A. S., Blott B. H. and Hames T. K. [1992], “Review of Neural Network Applications in
Medical Imaging and Signal Processing”,Medical and Biological Engineering and ComputingVol.
30, pp. 449-464.
[35] Fausett L. [1994], “Fundamentals of Neural Networks Architectures, Algorithms, and
Applications”, Prentice Hall, Englewood Cliffs, NJ.
[36] Anil K. Iain; Jianchang Mao, et K.M. Mobraddin: Artificial Neural Networks: A Tutorial 1996
[37] Widrow. B and Lehr, M. A., “ Thirty Years of Adaptive Neural Networks: Perceptron, MADALINE
, and Back propagation. Proc. of the IEEE, vol. 78, 1415-1442, 1990
[38] Rumelhart, D.E., Hinton G.E., and Williarns, R.J., Learning Internal Representations by Error
Propagation, Parallel Data Processing , vol. 1, chap 8 , The MIT Press, Cambridge, MA , 1986
[39] Specht, D.F., A General Regression Neural Network, IEEE Trans. Of Neural Networks, Vol. 2,
568 576, 1991
[40] Wasserman P.D., Advanced Methods in Neural Computing,VanNostrand Reinhold, New York, 1993
[41] Gibbons, R.J., et al., ACC/AHA Guidelines for Exercise Testing, A Report of American College of
Cardiology/American Heart Association Task Force on Practice Guidelines (Committee on Exercise
Testing), J. of the Am. Coll., of Cardiol., 30(1) 260-311, July 1997
AUTHORS
Atangana Romain was born in 1967 in Ngoumou-Cameroon. In 1997, he
graduated from the Higher technical Teacher’s Training College (ENSET)-
University of Douala with DIPET 1 (bachelor’s degree in electrical and
Electronics Engineering). In 2002 he obtained the DIPET 2 in the same
institution. He obtained a DEA (Master of Sciences with Thesis) at “Ecole
DoctoralePhysoique et Sciences de l’Ingénieur” in Electronics, Electrotechnic,
Automatic and Teledetection (EEAT) in 2007 at the University of Douala-
Cameroon. He is currently the Head of service of distance learning at the Higher
Teacher’s Training College (ENS) of Bertoua-University of Ngaoundéré. In parallel to his job, Mr.
Atangana Romain is following studies and research for a PhD thesis. His research interests are
telemedicine, signal and image processing, artificial intelligence and automation, medical physics, and
biomedical engineering. He is a member of “Unité de Recherche d’Automatique et
d’InformatiqueAppliquée (UR-AIA)” of the IUT-FV of Bandjoun of the University of Dschang-
Cameroon, and “Unité de Recherche de Matière Condensée et de Traitement de Signal (UR-MACETS)”
of the Faculty of Science of the University of Dschang-Cameroon.

32
Tchiotsop Daniel was born in 1965 in Tombel - Cameroon. He graduated in
Electromechanicalengineering from “Ecole Nationale Supérieure Polytechnique
(ENSP)” of Yaoundé-Cameroon in 1990, he obtained a MS degree in Solid
Physics in 1992 from the Faculty of Science of the University of Yaoundé I, a
MS degree in Electrical Engineering and Telecommunication in 2003 from
ENSP-Yaoundé and a PhD at INPL (“Institut National Polytechnique de
Lorraine”), Nancy-France, in 2007. Dr TCHIOTSOP teaches in the Department
of Electrical Engineering of “FOTSO Victor University Institute of Technology”
– University of Dschang since 1999 where he is actually the Head of Department. He is with
“Laboratoired’Automatique et d’InformatiqueAppliquée (LAIA)” where his main items of research
include Biomedical Engineering, Biomedical signal and image processing, Telemedicine and intelligent
systems.Dr TCHIOTSOP is partner with “Centre de Recherche enAutomatique de Nancy (CRAN)-
Université de Lorraine”, France, “Laboratoired’Electronique et du Traitement de Signal (LETS)-ENSP,
University of Yaoundé 1”, and “Unité de Recherche d’Automatique et d’InformatiqueAppliquée (UR-
AIA)” of the IUT-FV of Bandjoun of the University of Dschang-Cameroon.
Kenne Godpromesse was born in Mbouda-Cameroon in 1967. He obtaineda
B.S. degree in electromechanical engineering and a M.S. degree in mechanical
engineering from “Ecole Nationale Supérieure Polytechnique, Université de
Taoundé I”, Yaoundé-Cameroon in 1991 and 1994, respectively, and a PhD
degree in control theory from “Université Paris XI”, Orsay, France, in 2003.
Since 1996, he has been with the UniversityofDschang, Daschang-Cameroon,
where he is currently a Senior Lecturer. From 1996 to 2005, he has been the
Head of the Department of Electrical Engineering, Fotso Victor University
Institute of Technology,Bandjoun-Cameroon. He was a Guest Researcher at “Laboratoire des Signaux et
Systèmes (L2S)”, “Centre National de la Recherche Scientifique”, “L’Ecoled’Electricité (CNRS-
SUPELEC)”, “Université Paris XI”, and the Energy Department, SUPELEC, from September 2005 to
January 2006 and from June to July 2006. Since April 2006, he has been a founder of “Unité de
Recherche d’Automatique et d’InformatiqueAppliquée (UR-AIA)” of the IUT-FV of Bandjoun of the
University of Dschang-Cameroon. His current research interests include identification and control of
nonlinear systems using variable structure and neural network with applications in electromechanical
systems.
DjoufackNkengfack Laurent Chanel was born in 1991 in Nanga-Eboko-
Cameroon. In 2014, he obtained abachelor’s degree in physics option
Electronics, Electrotechnic and Automatic at the University of Dschang-
Cameroun. In 2016 he was graduating from the same institution with aMaster of
Sciences with thesis in Physics option Electronics, Electrotechnic and
Automatic. In 2018 he obtained a DIPET 2 in Electrical Engineering option
Electronics at the Higher Teacher’s Training College of Bambili-University of
Bamenda-Cameroon. He is currently a Technical High School teacher in
electronics engineering. In parallel to his job, Mr. Djoufack is following studies and research for a PhD
thesis. His research interests are telemedicine, signal and image processing, artificial intelligence, medical
physics, and biomedical engineering. He is a member of “Unité de Recherche d’Automatique et
d’InformatiqueAppliquée (UR-AIA)” of the IUT-FV of Bandjoun of the University of Dschang-
Cameroon, and “Unité de Recherche de Matière Condensée et de Traitement de Signal (UR-MACETS)”
of the Faculty of Science of the University of Dschang-Cameroon.

EEG SIGNAL CLASSIFICATION USING LDA AND MLP CLASSIFIER

More Related Content

Similar to EEG SIGNAL CLASSIFICATION USING LDA AND MLP CLASSIFIER (20)

More from hiij (20)

Recently uploaded (20)

EEG SIGNAL CLASSIFICATION USING LDA AND MLP CLASSIFIER