# Eeg Signals For Seizure Detection Biology Essay

Published:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Epileptic seizures are the outcome of the transient and sudden electrical disorder of the brain. The electroencephalogram is a therapeutic imaging method that measures the electrical activity in the brain. In this paper, Using discrete wavelet transform the EEG signals were decomposed by db1, db2 (Daubechies method) and Haar wavelet. GLCM and statistical features are used for extracting vital features from the decomposed EEG signal. The main principle of this study is to observe the performance of classifiers such as artificial neural network (ANN) and support vector machine (SVM) using wavelet coefficients for seizure detection.

Key words: EEG signal, Seizure, Epilepsy, GLCM, SVM, ANN, Wavelet, Accuracy

Introduction

To study brain function and neurological disorders, EEG is a complex signal plays an important role to get information about the electrical activity of the brain. Derivatives of epilepsy affect around 1% of the world population (Adeli, Zhou, & Datmehr, 2003). EEG records can provide helpful insight that keen on disorders of the brain activity. A seizure is a warning of epilepsy. Seizure is a paroxysmal event due to unbalanced excessive hypersynchronous neuronal discharge (Misra & Kalita, 2009). Epilepsy is known as a seizure disorder because the disarray of nervous system causing periodic loss of consciousness or convulsions. Special sensors (electrodes) are attached on the scalp to diagnose epilepsy and see what types of seizures are occurring (Suguna, 2012). There is demand for the automated seizure detection devices in the development of EEG recordings. Various pattern recognition methods for automated diagnosis have been adopted. In this paper, Using discrete wavelet transform the EEG signals both normal and seizure were decomposed at level 3 by db1,db2 (Daubechies method) and Haar wavelet associates. GLCM and statistical features are used for extracting vital features from the decomposed EEG signal. Six statistical features (mean, median, mode, standard deviation, skewness, and kurtosis) and GLCM features (contrast, correlation, energy, and homogeneity) are extracted for EEG signal classification. The performance of classifiers such as artificial neural network (ANN) and support vector machine (SVM) for seizure detection are analyzed. The entire procedure can be subdivided into a number of processing modules: preprocessing the EEG signal, decomposition of the signal using db1, db2 and haar wavelet, GLCM and statistical feature extraction and classification using ANN and SVM (Figure 1).

This paper is organized as follows. In Section 2, Literature reviews of EEG classification are discussed. Section 3 deals with materials and method. Section 4 gives evaluation of performance, followed by the conclusion at the Section 5.

Literature Reviews

(Tapan et al, 2012) proposed an idea which depends upon discrete wavelet packet transform (DWT) with mean, energy, standard deviation, entropy, kurtosis, skewness and entropy evaluation at each node of the decomposition tree pursued by application of probabilistic neural network (PNN). Approximation and details coefficients were developed using sixth-level DWT packet by decomposition of normal and epileptic EEG epochs. Their scheme obtained 100% accuracy for seizure detection. (Nicoletta et al, 2012) examined the use of Permutation Entropy (PE) as a feature for automated epileptic seizure detection. A Support Vector Machine (SVM) classifier was used for EEG classification. It was shown that average sensitivity of 94.38 and average specificity of 93.23 was achieved by using PE as a feature to characterize epileptic and seizure-free EEG. (Tapan et al, 2011) challenged to find the most useful wavelet function among the current members of the wavelet families for EEG signal analysis. Important features such as energy, entropy and standard deviation at different sub-bands were added for using the wavelet functions-Haar, Daubechies (orders 2-10), Coiflets (orders 1-10), and Biorthogonal (orders 1,1,2.4,3.5, and 4.4). The results acquired from PNN classifier were compared with Support Vector Machine (SVM). It was found that Coiflets 1 was the most suitable candidate among the wavelet families. (Abdulhamit Subasi et al, 2010)

Extracted statistical features from decomposed EEG signals into the frequency sub-bands using DWT to represent the distribution of wavelet coefficients. Principal components analysis (PCA), independent components analysis (ICA) and linear discriminate analysis (LDA) were employed to condense the dimension of data. Epileptic seizure was detected by using these features as the input to the SVM classifier. (Hasan Ocak 2009) analyzed EEG signals in two stages using approximate entropy (ApEn) and discrete wavelet transform (DWT) analysis. EEG signals were decomposed into approximation and detail coefficients using DWT in the first stage. In second stage, ApEn values of the approximation and detail coefficients were estimated for EEG signal classification. This scheme obtained 96% accuracy.

3. Materials and Method

3.1 Database

In this study, publicly available EEG databases were experimented and refer to Andrzejak et al, 2001 for further details. In our previous studies (Suguna 2012) we had been utilized this same database for artificial neural network (ANN) classifier and SVM classifier and concluded that SVM gives good performance classification results. In this proposed system, we have analyzed the performance of ANN and SVM classifier for the same set of database with wavelet features. Using the identical electrode placement scheme where five healthy volunteers were selected and EEG recording were carried out using sets A and B. Set A comprises those relaxed in an awaken state with eyes open and set B relaxed in an awaken state with eyes closed. Set C and D showed activity measured during seizure free intervals. Set E showed only seizure activity. Dataset (A and E) were used in this study. EEG recordings taken from set A and E are characterised in the following facts. The classifier ANN and SVM were trained using 50 datasets and tested using new set 20 epileptic and non epileptic data.

3.2 Discrete wavelet transform

The EEG signals can be measured as a superposition of dissimilar structures happening on different time scales at different times. One purpose of wavelet analysis is to separate and sort these underlying structures of different time scales. It is recognized that the wavelet transform (WT) is well suited to analyzing nonstationary signals. The property of time and frequency localization is known as compact support and is one of the most attractive features of the WT. In the study of EEG signal analysis, selection of appropriate wavelet and the number of decomposition levels are very important. In the present study, the number of decomposition levels was chosen to be 3. Thus, the EEG signals were decomposed into the details cD1- cD3 and one final approximation cA3 (Figure 2). Time-frequency signal processing algorithms such as discrete wavelet transform analysis are necessary to address different behavior of the EEG in order to describe it in the time and frequency domain. It represents a major advantage over spectral analysis. Hence DWT is well suited to locating transient events. Such transient events as spikes can occur during epileptic seizures (Adeli, Zhou, & Dadmehr, 2003; Subasi 2007).

Maximum efficiency of different types of wavelet is selected and performed research for seizure detection. The large amount of identified wavelet families and functions offers a rich space in which to investigate for a wavelet which will professionally represent a signal of interest in a variety of applications. Biorthogonal, Coiflet, Haar, Symmlet, and Daubechies wavelets are belong to wavelet families. There is no fixed way to decide a certain wavelet. The selections of the wavelet function depend on the application. The Daubechies algorithm is theoretically complex and has a higher computational overhead. The Haar wavelet algorithm has the benefit of being simple to calculate and easy to understand. Selecting a wavelet function which closely matches the signal to be processed is of utmost important in wavelet applications. The even feature of the Daubechies wavelet of order 2 is more suitable for detecting changes of the EEG signals. The approximation coefficients of normal and abnormal subjects using Haar wavelet are shown in the following Figure 3.

3.3 Statistical Feature Extraction

Feature extraction acting a significant role in pulling out unique patterns from the original data for consistent classification. Six statistical features (mean, median, mode, standard deviation, skewness and kurtosis) of each wavelet are used as the valuable parameters for the representation of the characteristics of the original EEG signals. The sample statistical features for normal and abnormal subjects are shown in the following table 1.

3.4 GLCM Features

The EEG signal features are extracted using GLCM (Gray level Cooccurrence matrix). A Cooccurrence matrix C is defined over n*m and image I parameterized by an offset ( ) as

The sample GLCM features for normal and abnormal subjects are shown in the following table 2.

3.5 Artificial Neural Network Classifier

Artificial neural networks are developed based on brain structure. Artificial neural networks can recognize patterns, manage data and learn just like brain. They are made up of artificial neurons which realize the spirit of biological neurons. It accepts a number of inputs. Each input comes via connection, which is called synapses and which has a weight. A neuron also has a threshold value. The neuron is stimulated if sum of the weights is greater than this threshold value. The activation signal creates the output of the neuron which will be the result of the problem or can be judged an input for another neuron. A number of neurons forms together to create an artificial neural network which is arranged on layers. A network has to contain an input layer (which takes the values of outside variables) and an output layer (the predictions or the result). Inputs and outputs correspond to sensory and motor nerves from human body. There also can be hidden layer(s) of neurons, which play an internal role in the network. All these neurons are connected together. Selected GLCM and statistical features from the decomposed EEG signals are used to train and test the network for classification of normal and abnormal patterns.

3.6 SVM Classifier

SVM has been used during current years as an alternative to ANN. It attains enhanced generalization due to structural risk minimization (SRM). SVM is linear, but it can be used for non linear data by using kernel function. Basic SVM is a two class classifier. However, with some modification, multiclass classifier can be obtained. Classifying data is an important job in machine learning. In EEG classification problems, we are given n experiments {(x1, y1),â€¦. (xn,yn)} This is called the training set. Where xi is a vector and yi is a binary class label Â± 1. For this given data points each data pointed belong to one of the two classes. The main intend of the SVM algorithm is to choose which category a new data position will be in. The SVM classifier was trained using MATLAB. The function svmclassify classifies each row of data in sample using the information in the support vector machine classifier structure svmstruct, created using svmtrain function (Suguna 2012). In this study, The SVM is applied to the GLCM and Statistical features for EEG signal classification. The pseudo code for entire process is as follows;

Step 1: Getting EEG signal from the source

Step 2: After pre-processing, EEG signals are decomposed using db1, db2 and haar wavelet.

Step 3: Extracting GLCM/Statistical/hybrid features from the decomposed signal.

Step 4: These features were computed and classified the normal and abnormal subjects using the classifier ANN and SVM.

Step 5: The performance of classifiers are analyzed using confusion matrix.

4. Evaluation of performance

A value of "-1" was used when the experimental investigation indicated a normal EEG pattern and "+1" for epileptic seizure. Prediction success of the classifier may be evaluated by examining the confusion matrix. In order to analyse the output data obtained from the application, TPR (true positive ratio) and TNR (true negative ratio) are calculated by using confusion matrix. Sensitivity, specificity and total classification accuracy are calculated by the following formula.

Sensitivity = TPR =

Specificity = TNR =

Accuracy =

Figure 4 shows; using haar wavelet SVM classifies the statistical features of approximation signal and SVM classifies the GLCM features of approximation signal using db1 are shown in Figure 5.

From our previous studies table 3, we have extracted GLCM features (contrast, correlation, energy, homogeneity) without wavelet decomposition from the EEG signals and we used ANN and SVM classifier to identify the normal and abnormal subjects. We have achieved 85% accuracy in ANN and 90% accuracy in SVM. So we come to the conclusion and find that SVM is the best method for EEG signal classification. Figure 6 proves the analysis of ANN and SVM classifier without wavelet. Even though SVM is the best method which we had considered, it was analyzed without wavelet coefficients. Since wavelet transform is compatible for non stationary signals, we have analyzed the EEG signals with wavelet decomposition in this proposed system. Here we have used GLCM features, six statistical (mean, standard deviation, median, mode, skewness, kurtosis) and hybrid features for different kinds of wavelet members (db1, db2, haar). We have studied the comparison of the performance yield by the classifiers ANN and SVM.

Table 4 shows the comparison of classifier ANN and SVM results using db1. The SVM classified the normal and abnormal subjects with the accuracy of 80% compared to ANN accuracy of 70% with GLCM features using db1 wavelet at approximation coefficient (cA3).Regarding statistical features ANN achieved 95% whereas SVM achieved 100% accuracy. Presentation of hybrid features reached 100% accuracy both in ANN and SVM classifier. So performance of these classifiers in approximation coefficient is efficient. Similarly we can analyze the classification results which are articulated in terms of sensitivity, specificity and accuracy for detail coefficients also. When GLCM features acts as an input to the classifier, in the detail coefficient (cD3) ANN achieved 85% whereas SVM attained only 80%. When we are considering detail coefficient level 2 (cD2) ANN attained 75% whereas SVM achieved 80% accuracy. In detail coefficient level 1 (cD1) ANN obtained 95% whereas SVM reached 100% accuracy. When expressing the ideas of GLCM features the above mentioned accuracies are good both in ANN and SVM. When we are considering Statistical features SVM yields 100% in approximation coefficient (cA3) as well as in detail coefficient (cD1) whereas ANN achieved maximum 95%. While discussing about hybrid (GLCM and Statistical) features SVM as well as ANN reached 100% accuracy in approximation coefficient (cA3). From the analysis of classification results using db1 wavelet we come to the decision that hybrid features deserves good results for EEG signal classification. Figure 7 and 8 are the diagrammatic representation for the analysis of ANN classifier using db1 wavelet and analysis of SVM classifier respectively.

Table 5 shows the comparison of classifier ANN and SVM results using haar wavelet. The SVM and ANN classifier categorized the normal and abnormal subjects with the accuracy of 80% with GLCM features using haar wavelet at approximation coefficient (cA3).Regarding statistical features ANN achieved 95% whereas SVM achieved 100% accuracy. Presentation of hybrid features reached 100% accuracy both in ANN and SVM classifier. So performances of these classifiers in approximation coefficient are efficient. Similarly we can analyze the classification results which are expressed in terms of sensitivity, specificity and accuracy for detail coefficients also. When GLCM features acts as an input to the classifier, in the detail coefficient (cD3) ANN achieved 100% whereas SVM attained only 80%. When we are considering detail coefficient level 2 (cD2) ANN attained 75% whereas SVM achieved 80% accuracy. In detail coefficient level 1 (cD1) ANN obtained 95% whereas SVM reached 100% accuracy. When communicating the ideas of GLCM features the above mentioned accuracies are good both in ANN and SVM. When we are considering Statistical features SVM yields 100% in approximation coefficient (cA3) as well as in detail coefficient (cD3) whereas ANN achieved maximum 95%. While discussing about hybrid (GLCM and Statistical) features SVM as well as ANN reached 100% accuracy in approximation coefficient (cA3). From the analysis of classification outcomes using haar wavelet we come to the result that hybrid features and SVM classifier deserves good results for EEG signal classification. Figure 9 and 10 are the diagrammatic representation for the analysis of ANN classifier using haar wavelet and analysis of SVM classifier respectively.

Table 6 shows the comparison of classifier ANN and SVM results using db2. The SVM classified the normal and abnormal subjects with the accuracy of 75% compared to ANN accuracy of 65% with GLCM features using db2 wavelet at approximation coefficient (cA3).Regarding statistical features and hybrid features both ANN achieved 100% accuracy. So performances of these classifiers in approximation coefficient are efficient. Similarly we can analyze the classification results for detail coefficients also. When GLCM features acts as an input to the classifier, in the detail coefficient (cD3) ANN achieved 90% whereas SVM attained only 95%. When we are considering detail coefficient level 2 (cD2) ANN attained 85% whereas SVM achieved 90% accuracy. In detail coefficient level 1 (cD1) ANN obtained 80% whereas SVM reached 85% accuracy. When conveying the ideas of GLCM features the above mentioned accuracies are good both in ANN and SVM. When we are considering Statistical features SVM yields 100% in approximation coefficient (cA3) as well as in detail coefficient (cD3) whereas ANN also achieved 100% at approximation level. While discussing about hybrid (GLCM and Statistical) features SVM as well as ANN reached 100% accuracy in approximation coefficient (cA3). From the analysis of classification results using db2 wavelet we come to the conclusion that hybrid features and SVM using db2 wavelet deserves the best combination for EEG signal classification. Figure 11 and 12 are the diagrammatic representation for the analysis of ANN classifier using db1 wavelet and analysis of SVM classifier respectively. The result for seizure detection shows the improved performance with the earlier work with the same dataset. This study proves the importance of the wavelet decomposition for non stationary EEG signals.

5. Conclusion

A methodical analysis of EEG for seizure detection using wavelet based features has been presented in our proposed system. As EEG is a nonstationary signal the discrete wavelet transform yields good outcome. After wavelet decomposition at level 3 using db1, db2 and haar wavelet, six statistical features and GLCM features are well analyzed over the wavelet coefficients at each level. Classification has done using ANN and SVM. The comparisons of classification results has been analyzed and found that SVM classifier with hybrid features using db2 wavelet functions deserves the best combination for automated seizure detection. In near future, wavelet packet entropy feature will be applied to the increased data sets and other linear classifier can be analyzed for automated seizure detection.

EEG Signal

Preprocessing

Decomposition (dwt)

## (

Feature Extraction

Classifier

SVM

Neural Network

Seizure Prediction

DB1

Haar

DB2

Figure 1: Modules of Proposed System

Signal

cA2

cD1

cA1

cA3

cD3

cD2

Figure 2: Decomposition of wavelet transform

Figure 3: Approximation Coefficients Signal of Normal and Abnormal Subjects

Table 1: Statistical features for approximation signal using db1

Features

Subjects

mean

Std Deviation

median

mode

skewness

kurtosis

Normal

S1

19.6184

99.7745

20.8597

-43.4871

-0.2345

3.5839

S2

-148.4786

107.3429

-145.6640

-253.1442

-0.2205

3.5141

S3

36.2699

110.1132

36.7696

83.4386

0.0737

3.2320

Abnormal

S1

0.1352

1.0270

0.3702

1.0437

-0.0008

0.00031

S2

0.1070

1.0708

0.1743

0.3338

-0.0003

0.0028

S3

0.0905

0.6954

0.1761

-0.0665

-0.0002

0.0030

Table 2: GLCM features for approximation signal using db1

Features

Subjects

Contrast

Correlation

Energy

homogeneity

Normal

S1

24.9502

-0.0528

0.2612

0.5506

S2

8.1688

-0.0203

0.6937

0.8525

S3

24.0758

-0.0658

0.2891

0.5701

Abnormal

S1

20.8939

0.0866

0.2886

0.6269

S2

25.2424

-0.0468

0.2581

0.5492

S3

22.0823

0.0690

0.2587

0.5997

Figure 4: SVM classifies the statistical features of approximation signal using haar wavelet

Figure 5: SVM classifies the GLCM features of approximation signal using db1 wavelet

Table 3: performances of classifiers ANN and SVM without wavelet decomposition

Measures

ANN Values (%)

SVM Values (%)

Sensitivity

100

90

Specificity

76

90

Accuracy

85

90

Figure 6: Analysis of ANN and SVM classifier without wavelet decomposition

Wavelet

Co-efficient

Extracted

Features

Neural Network

SVM

Sensitivity

## %

Specificity

## %

Accuracy

## %

Sensitivity

## %

Specificity

## %

Accuracy

## %

db1

cA3

GLCM

70

70

70

80

80

80

Statistical

90.9

100

95

100

100

100

GLCM and Statistical

100

100

100

100

100

100

cD3

GLCM

88.8

81.8

85

80

80

80

Statistical

90.9

100

95

100

100

100

GLCM and Statistical

90.9

100

95

90.9

100

95

cD2

GLCM

77.7

72.7

75

80

80

80

Statistical

76.9

100

85

90.9

100

95

GLCM and Statistical

90.9

100

95

100

100

100

cD1

GLCM

100

90.9

95

100

100

100

Statistical

76.9

100

85

71.4

100

80

GLCM and Statistical

83.3

100

90

83.3

100

90

Table 4: Comparison of Classification results using db1

Table 5: Comparison of Classification results using Haar

haar

Wavelet

Co-efficient

Extracted

Features

Neural Network

SVM

Sensitivity

## %

Specificity

## %

Accuracy

## %

Sensitivity

## %

Specificity

## %

Accuracy

## %

cA3

GLCM

75

87.5

80

80

80

80

Statistical

90.9

100

95

100

100

100

GLCM and Statistical

100

100

100

100

100

100

cD3

GLCM

100

100

100

80

80

80

Statistical

90.9

100

95

100

100

100

GLCM and Statistical

90.9

100

95

90.9

100

95

cD2

GLCM

77.7

72.7

75

80

80

80

Statistical

76.9

100

85

90.9

100

95

GLCM and Statistical

83.3

100

90

100

100

100

cD1

GLCM

100

90.9

95

100

100

100

Statistical

71.4

100

80

71.4

100

80

GLCM and Statistical

83.3

100

90

83.3

100

90

Table 6: Comparison of Classification results using db2

db2

Wavelet

Co-efficient

Extracted

Features

Neural Network

SVM

Sensitivity

## %

Specificity

## %

Accuracy

## %

Sensitivity

## %

Specificity

## %

Accuracy

## %

cA3

GLCM

71.4

61.5

65

72.7

77.7

75

Statistical

100

100

100

100

100

100

GLCM and Statistical

100

100

100

100

100

100

cD3

GLCM

90

90

90

100

90.9

95

Statistical

100

90.9

95

100

100

100

GLCM and Statistical

100

100

100

100

90.9

95

cD2

GLCM

88.8

81.8

85

90

90

90

Statistical

62.5

100

70

90.9

100

95

GLCM and Statistical

90.9

100

95

90.9

100

95

cD1

GLCM

87.5

75

80

88.8

81.8

85

Statistical

90.9

100

95

90.9

100

95

GLCM and Statistical

75

87.5

80

90.9

100

95