Epilepsy EEG Classification Based on Convolution Support Vector Machine_Qi Xin.pdf

Hubble

154

8页

3次

2023-01-30

免费下载

Printed in the United States of America

RESEARCH ARTICLE

Journal of Medical Imaging and

Health Informatics

Vol. 11, 25–32, 2021

Epilepsy EEG Classiﬁcation Based on Convolution

Support Vector Machine

Qi Xin

, Shaohai Hu

1 ∗

, Shuaiqi Liu

2 ∗

, Xiaole Ma

, Hui Lv

, and Yu-Dong Zhang

4 ∗

Institute of Information Science, Beijing Jiaotong Universi ty, Beijing 100044, China

College of Electronic and Information Engineering, Hebei University, Baoding Hebei 071002, China

Beagledata Technology (Beijing) Co., Ltd., Beijing, 100089, China

Department of Informatics, University of Leicester, Leicester, LE1 7RH, UK

Clinical Electroencephalogram (EEG) data is of great signiﬁcance to realize automatable detection, recognition

and diagnosis to reduce the valuable diagnosis time. To make a classiﬁcation of epilepsy, we constructed

convolution support vector machine (CSVM) by integrating the advantages of convolutional neural networks

(CNN) and support vector machine (SVM). To distinguish the focal and non-focal epilepsy EEG signals, we

ﬁrstly reduced the dimensionality of EEG signals by using principal component analysis (PCA). After that, we

classiﬁed the epilepsy EEG signals by the CSVM. The accuracy, sensitivity and speciﬁcity of our method reach

up to 99.56%, 99.72% and 99.52% respectively, which are competitive than the widely acceptable algor ithms.

The proposed automatic end to end epilepsy EEG signals classiﬁcation algorithm provides a better reference

for clinical epilepsy diagnosis.

Keywords: EEG Signal Classiﬁcation, PCA, CNN, SVM.

1. INTRODUCTION

The sudden unexpected electro-discharge of brain neurons will

bring to transience brain dysfunction, which is the mechanism

of epilepsy (a chronic illness) [1]. Epilepsy is caused by many

factors, such as genetic factors, cerebral diseases and system-

atic diseases. Epilepsy is the second most common destructive

nervous system disease [2], which affecting 0.4%∼1% of world

population [2]. One-third of epilepsy patients are not a vailable

to hav e appropriate medical treatment, and the treatment of the

remaining two-thirds varies according to the epilepsy status. Dur-

ing an epileptic seizure, the patient may have a slight temporal

clouding of consciousness or uncontrollable convulsions.

There are many clinical diagnostic methods for various dis-

eases, such as Magnetic Resonance Imaging for tumor, Diffu-

sion Tensor Imaging for neurodegenerative disorders, Computed

Tomography for orthopedic diseases and EEG [3]. The EEG-

based methods are mainly used for epilepsy detection in gen-

eral [4]. It has been demonstrated that there is signiﬁcant

difference between the epileptic EEG and the normal one. The

cerebration of epilepsy includes ictal and inter-ictal periods, and

the change of EEG signals is obvious during the period of cyclic

transition, so EEG monitoring is the key measure to distinguish

whether epileptic seizures or not [5]. Figure 1 gives the general

view of brain electrical activities of normal and epileptic seizures.

∗

Authors to whom correspondence should be addressed.

Clinically, epileptic patients are assessed with neurosurgical

treatment by e xcising the epileptic seizures brain regions. EEG

signal can be recorded from brain surface and even the deeper

brain structure in the process of some epileptic diagnosis. The

dynamic characteristics of these EEG records in ictal and inter-

ictal stages both provide valuable information for epileptogenic

focus location. At present, the epileptic diagnosis is carried out

by comprehensive analysis of the patient’s EEG, as well as symp-

toms, signs and other experimental results. However, manual

observation and detection of long-term EEG signals are complex

and time-consuming. Prolonged observ ation may cause fatigue of

medical personnel, which may easily lead to inaccurate manual

test results. Besides, professional experience is quite important in

manual observation [6], which demands the professional knowl-

edge of epilepsy for doctors. Therefore, the development of auto-

matic epileptic EEG signal classiﬁcation algorithm is of great

signiﬁcance. Recent research shows inter-disciplinary approach

can provide worthy diagnostic information for location of epilep-

tic focus under interictal and ictal period [7–9].

Since the 1960s, the automatic classiﬁcation of epilepsy EEG

has attracted extensive research interest and a variety of auto-

matic detection methods in epilepsy are appeared. For instance,

the template matching based method proposed in Ref. [10] for

epilepsy identiﬁcation taking the selected typical epilepsy signals

as a template and matching the detected waveform to identify

the disease. However, the diversity of EEG signals leads to great

J. Med. Imaging Health Inf. Vol. 11, No. 1, 2021 2156-7018/2021/11/025/008 doi:10.1166/jmihi.2021.3259 25

RESEARCH ARTICLE J. Med. Imaging Health Inf. 11, 25–32, 2021

Fig. 1. The general view of brain electrical activities of normal and epileptic

seizures.

difﬁculties in samples selection. Acharya distinguishes normal

EEG and epileptic EEG through nonlinear ﬁltering [11]. Ho w-

ever, the false detection rate is higher if many false differences,

i.e. e y e movement, myoelectric, are existed in the EEG signals.

A frequency-domain and digital ﬁlter based EEG classiﬁcation

algorithm is recommended to generate EEG reports automati-

cally in Ref. [12]. On the basis of the ﬂuctuation of the fre-

quency range relative to the background employing wavelet, a

sensitive but less mis-detected method for epilepsy classiﬁcation

is proposed [13]. Many researchers detect epilepsy by extracting

nonlinear EEG features. For example, in Ref. [14], Srinivasan

et al. put forward an epileptic EEG signal classiﬁcation algo-

rithm by combining artiﬁcial neural networks and approximate

entropy. In Ref. [15], the authors presented an epilepsy classiﬁ-

cation method based on Lyapunov index. In Ref. [16], Acharya

et al. used four different kinds of entropy to classify epileptic

seizure and achie ved good results. In Ref. [17], fuzzy entropy

and distributed entropy were used for epilepsy classiﬁcation. In

Ref. [18], a new single feature called Tsallis entropy was used

for epilepsy classiﬁcation, which greatly shortened t he detection

time. The rapid growth of expert system and fuzzy technology

has emphasized the ﬁeld of epilepsy classiﬁcation [18–20].

The synchronous mode of epilepsy is usually nonlinear, diverse

and uncertain, and there is difﬁculty in identifying the period of

epilepsy based on linear method. Therefore, the nonlinear adap-

tive pattern recognition is imperative. Deep learning can provide

nonlinear processing, adaptability and strong fault-tolerant for

signals. As deep learning has strong ability of automatic features

learning, it performs well in pattern recognition. As a deep learn-

ing model, CNN [21] is widely used in various signal processing

and image processing research ﬁelds. Similar to BP (Back Prop-

agation), the forward propagation is employed by CNN model to

calculate the output value, and it also uses back propagation to

adjust the weight and bias of the network. However, the biggest

difference between CNN and BP is that the neural units between

adjacent layers in CNN are partially connected, which are not

fully connected. That is, the sensory area of a certain neural

unit comes from the upper part of the neural units, instead of

being connected to all the neural units as BP does. Besides, as

a deep neural network structure, CNN does not have the same

incomputable problem as the traditional deep neural network.

The topological structure of data and the unkno wn parameters in

network can be solved by optimizing the network structure based

on the back-propagation algorithm. Therefore, CNN is widely

applied in various ﬁelds, such as signal denoising [22], signal

classiﬁcation [23], machine health diagnosis [24] and cancer pre-

diction [25]. Recently, the classiﬁcation of epilepsy based on

deep learning has been preliminarily entered on the scene. For

example, Ullah recommends the pyramid one-dimensional CNN

model to predict the types of EEG in epilepsy test data, and the

model simpliﬁes the parameters of CNN.

Although CNN can be well applied for signal classiﬁcation,

the training of CNN is time-consuming for large signal data.

To solve the problem, we ﬁrstly carry out principal component

analysis of EEG data and pick up the essential components of

EEG signals as the training and recognition features to accelerate

the CNN training process. Although the massive data guaran-

tees a good classiﬁcation effect for CNN model, the CNN-based

classiﬁcation algorithm does not get the desired effects if the

sample size is insufﬁcient. Based on the structural risk mini-

mization, SVM can achieve better classiﬁcation results with the

insufﬁcient number. SVM is a new machine learning method to

ﬁnd the maximum classiﬁcation interval on account of the theory

of statistical learning [27]. SVM is a kind of machine learning

methods accorded with the structural risk minimization criterion.

It can obtain small errors after the training of a little training

set, and SVM can keep small errors to another stand-alone test

set. For classiﬁcation, SVM gives the category of unknown sam-

ples by the decision surface of different classes in the region.

In Ref. [28], Siuly et al. presented an EEG signal classiﬁca-

tion algorithm based on SVM. This algorithm applied LS-SVM

to class features that extracted by clustering in EEG data. This

algorithm performs well with the price of time-consuming. As a

commonly used supervised machine learning method in the ﬁeld

of classiﬁcation, SVM [29–31] can screen a few most effective

training samples for the prediction task from high-dimensional

data and it has a strong generalization ability, which can be used

for nonlinear classiﬁcation.

To combine the advantages of CNN and SVM, a convolu-

tion support vector machine named CSVM is constructed for

the classiﬁcation of focal and nonfocal in epilepsy EEG sig-

nals. CSVM is constructed by removing CNN’s soft-max layer

and directly cascading SVM. CSVM can classify epilepsy EEG

signals through CNN to extract the features of the principal

components after PCA. CSVM not only effectively increases

the accuracy of classiﬁcation of epileptic EEG signals, but also

reduces the time of training and recognition, which is valuable

in clinical application.

2. THEORETICAL BASIS

2.1. Principal Component Analysis

In many research areas, researchers need to look for patterns in

the analyzed mass of data. What’s worse, the data to be analyzed

usually has multiple variables. While bringing a large amount of

information, the mass data also increases the workload of data

collection and processing, and the related variables also bring

challenges to the data processing. As the number of variables

decreases, the information contained in the data decreases to

some extent, we must ﬁnd a feasible method to maintain the

main information and reduce the loss of information contained in

the original mass data. While reducing the index to be analyzed,

and PCA is a method which can reduce data dimensionality by

using eigenvectors based on multivariate statistical analysis. The

high-dimension eigenvectors is transformed into low-dimension