Performance Analysis of Fuzzy Multilayer Support Vector Machine for Epileptic Seizure Disorder Classification using Auto Regression Features

1.1. EEG Epileptic Seizure Database

The databases (EEG signal) for this research were collected from Karunya University. The performance results are compared with K-means, Navies Bayesian and Neural Network with respect to the parametric method. The features were extracted by applying feature extraction methods and the extracted features were classified by implementing the Elman Neural Network.

The EEG data utilized in this research was obtained utilizing 10-20 electrodes, specified by the standard international framework. The EEG data was downloaded from the database discovered in the link http://karunya.edu/ research/EEGdatabase/public/view_all.php. The data were recorded from 18 channels (2 periocular electrodes and 16 scalp electrodes, with reference to right and left mastoid) at a sampling rate of 256 Hz with analog passband of 0.01 to 100 Hz. Before setting the electrodes, the scalp was arranged in order to contact impedance below 5kΩ. EEG data generation had a 10-second term, involving 2560 data points, all having 4-ms duration. This database contains data on clinical status of the patients [3].

1.2. Selection of Leads

The recording of the EEG signal consisted of 32 channels that were acquired by placing metal electrodes on the scalp region. The position of the electrodes was indicated in the standard 10-20 electrode placement system. For this study, 8 channels were selected from Frontal Pole to Occipital region covering both the hemispheres of the cerebral cortex. During the onset of epileptic seizures, the cerebral functions of the human brain showed less randomness. This was captured by the EEG electrodes and transmitted to the detection system for processing. Fig (2) exhibits the 10-20 EEG lead placement system that is used for data acquisition [3-5].

1.3. Pre-processing

The pre-processing stage separated the input EEG data into different multichannel signals. The 8 signals corresponding to FP1-F3, F3-C3, C3-P3, P3-O1, FP2-F4, F4-C4, C4-P4 and P4-O2 were used for the detection of epilepsy in this study . The selected lead locations are highlighted in Fig. (2). The channels were sampled at 256 Hz and are noise-free. The data sets used for training and testing the classifier had the same time duration of 60 sec [3-5].

An extensive analysis of the brain signal could expose the decrease in the randomness of the EEG signal as the count of neurons comprising brain functions decreased during epileptic seizures. This characteristic nature of Epilepsy is used as an important classification rule in the system [3-5].

1.4. Feature Extraction

Generally, feature extraction is the process inspiring important features from a set of signals that inclines suitable classification and disease analysis. A measure which indicates the level to which the ups and downs of a wave vary on an average out of the mean voltage was said to be a SD. Since decomposing the signal, some important features could be extricated utilizing the SD. For dissecting the statistical feature of a signal, the signal's mean must be calculated. It can likewise be stated that the value of mean is the average estimation of a signal.

The feature extraction procedure is used to discover the average accuracy of the framework. The accompanying coefficients like Autoregression (AR) Burg, AR Covariance, AR Yule-Walker technique, AR Modified Covariance, Linear Prediction Coefficient, and Levinson Durbin Recursion were chosen for validation of the extraction procedure. It reduces the EEG structures and modules dependent on the training procedure.

3.1. Cascade Forward Backpropagation Neural Network (CFBNN)

CFBNN is similar to Backpropagation Artificial Neural Network (BPANN) which uses the Back Propagation algorithm for updating weights. CFBNN has a weight connection from the input to every layer and from every layer to the following layers. The network activation flow was one-way only, from the input to the output layer passing over the hidden layer. Fig. (3) presents a three-layer network that has links from layer 1 to layer 2, layer 2 to layer 3, and layer 1 to layer 3. More connections may enhance the speed at which the network recognizes the required relationship. Back propagation algorithm is used for updating weights. The errors propagate backward from the output nodes to the input nodes [10].

Fig (5) demonstrates the architecture of CFBNN. In this architecture, the neurons in each layer relate to the neurons in the past layer. The layers of the cascade neural networks are active in quality. With regard to cascade networks, with the support of one neuron, learning is initiated amid the training stage, the learning algorithm includes new neurons subsequently making a multilayer architecture. If the training error diminishes, it suggests that the aggregate number of hidden layer of neurons increments in a well-ordered manner. The development of the network is controlled by the learning algorithm. For the enhancement of the loads in the network, two training algorithm, Levenberg-Marquardt and Bayesian control back propagation algorithms are utilized. It can result in a quick learning rate when the availability is controlled inside the network [11 – 15].

3.2. Fuzzy Multilayer Support Vector Machine Classification

Data classification is the way toward classifying and ordering the data into various types, classes, and different structures. To create a different objective, the classification and partitioning are facilitated by the classification of data according to the points of data. To separate whether the subject is epileptic or not, classification of data has been conducted . For accumulating data and various strategies, data classification is an important approach. To classify the epileptic seizure in this unit, FMSVM technique has been used. Fuzzy rule optimizes the MSVM parameters to improve the MSVM execution.

For performing different operations, feature extraction and pre-processing were conducted in the EEG signals. Uncontaminated non-cerebral artifacts are the significant issue obtained in clean information on brain activity. A hindrance was created during the task of decision-making , while occurrences of artifacts were detected in the EEG signals. Eyeblink and body movements or certain sources like electrical impedances were created by the subject during the process of data acquisition.

The proposed three-class classification of epileptic seizures is assigned a class label of yϵ(0,−1,+1), where 0, +1, and -1 represents seizure disorders for an input feature vector xϵR_N. Consider an input-output training pair { (x₁, y₁), …, (i_n, y_n) }, where vectors x₁, x₂, …, x_n are input vectors and y₁, y₂, …, y_n are the required output vectors. The issue is to identify a classifier f(x) which classifies the input samples.

These features were an input pattern and labeled as +1 for seizure presence and –1 for seizure absence. Collectively(x, y) forms an input-output pair, where x was the input feature and y was the output.

The fuzzy parameters utilized as a part of this work are necessary for a calculated relapse membership function. FMSVM implements a fuzzy membership capacity for every input data of SVM. SVM is receptive to exceptions and noises in the preparation of information that does not have a place with the three classes. Every data in the training dataset is doled out with membership and the total of the deviations weighted by its membership.

By allocating a low membership value to an outlier, the error in the training method could be decreased. The use of FMSVM is beneficial as it has less membership value and would be allocated to data points that differ from the classes. The error in training is decreased by achieving better generalization capacity. The membership rate for the membership capacities was appointed a value in the interval of [0, 1]. In the FMSVM classifier, the regularization parameter was set as C = 10, σ2 = 1 for kernel. The trial outcomes show the solidity of the present method.

FMSVMM figures out how to evaluate the objective capacity utilizing the extricated vector as input. For registering the hidden layer portrayal f(x|θ) of input vector x, these works make utilization of:

Which is iteratively used by each SVM to compute the element f(x|θ)a. The target values for hidden-layer features were between -1 and 1, therefore if some target output is larger than 1 for a feature, the target value should be set at 1. To allow the hidden-layer SVMs to extract different features, breaking symmetry was necessary. For this, this study randomly initializes the trainable parameters in each hidden-layer SVM. However, a better way to initialize the hidden-layer SVM is to let them train on different perturbed versions of the target outputs. Therefore, a dataset should be constructed (x_i, y_i + γ a i), with γ a i random value ϵ [−γ, γ] for the hidden-layer SVM Sa, where γ is another meta parameter.

FMSVM have an excellent potential to be connected in programmed identification of variations from the seizures by lessening the misclassification rate. In the analysis of research, obtaining an accurate presumed area in seizure is a vital issue. The relating execution indicates that FMSVM is higher in terms of accuracy, specificity, sensitivity, and misclassification rate. The outcomes acquired from testing exhibited which fuzzy-based SVM classifier produced the best execution, dominating regular SVM [16].

3.3. Autoregression Features

The autoregressive model was utilized to decrease the least square model and errors of prediction. This module is beneficial because it stays constant while processing the signal. The contribution to this module is in the structure of a column vector. Its parameters might be important as far as both various co-efficient and reflection coefficients are concerned. The AR module gives an elective method for examining the EEG spectral properties estimation [17-20]. Related on the discrete linear stochastic procedure, it is stated as,

The errors are stated as, e_t = y_t - µ - ψ₁ e_t-1 - ψ₂ e_t-2 - ··· .

Assume the stationarity model which holds for e_t should hold true for e_t-1, then e_t-1 = y_t-1 - µ - ψ₁ e_t-2 - ψ₂ e_t-3 - ··· .

Finally, substitute the model for e_t-1 into the model for y_t

(1)

Where, y₁, y₂, ···, y_n are the observations with a joint density Pr(y₁, y₂, ···, y_n). e_t are the errors with respect to time.

a) AR Covariance

The parameter γ_j was called as the autocovariance at lag j, combining all results together, it would be

(2)

(3)

The covariance matrices, V(y) were symmetric. If E(y_t) does not base on t then it must not be a stationary series, so we can normally anticipate detecting the series in the neighborhood of µ.

(4)

If γ_j > 0, we could anticipate when a higher than normal observation could be followed by another higher than normal observation. We could normalize the covariance by describing the autocorrelation,

(5)

As usual, ρ₀ = 1. The structure of the autocorrelations would significantly support us in knowing the conduct of the series, y [17-20].

b) AR Yule-Walker

It calculates the AR parameters by shaping a biased measure of the signal's autocorrelation and comprehending the least square minimization of the forward prediction error. Here, the procedure linearly relies on the amplitude of a signal at the provided duration. The amplitude was acquired by adding various amplitudes of the past samples and error of estimation. The order for the filter linearly relies upon the count of AR coefficients. The modeling degree (p) consistently utilizes the Akaike Information Criteria (AIC) [17-20].

Generally, an AR model of order p could be written as

(6)

The autocorrelations and the ϕ_i were incidental to each other, known as the Yule-Walker Equations:

(7)

Which could be utilized to measure values.

The Yule-Walker AR model is estimated by reducing the value of the prediction error power [17-20].

c) AR Modified Covariance

The AR modified covariance was utilized to evaluate the PSD of an EEG input signal. The principal reason for this research is to reduce the forward and backward anticipation errors . At last, the evaluation order parameter should be not exactly, or equivalent to, 2:3 of the input vector length to settle the output. This procedure was completely characterized by a direct combination of past results and driving noise. It evaluates the P coefficients, where P is the model, by reducing the forward and backward prediction errors at all squares sense:

(8)

(9)

Where for the data length was N and was the AR coefficient of the k^th term [17-20].

d) Levinson Durbin Recursion

It is a normal algorithm that is simple to solve; here, the model for k =1 and k +1 coefficients estimated issues. The initial step completed in LDR is to reduce the error. At that point, the input vector and error vector are figured. The k values from o to m are evaluated [17-20].

e) Linear Prediction Coefficient

In the linear prediction auto-correlation technique, the linear prediction coefficientsare calculated from the Bartlett-window-biased autocorrelation function [17-20].