Umut Orhan

Short Communication: EEG signals classification using the K-means clustering and a multilayer perceptron neural network model

We introduced a multilayer perceptron neural network (MLPNN) based classification model as a diagnostic decision support mechanism in the epilepsy treatment. EEG signals were decomposed into frequency sub-bands using discrete wavelet transform (DWT). The wavelet coefficients were clustered using the K-means algorithm for each frequency sub-band. The probability distributions were computed according to distribution of wavelet coefficients to the clusters, and then used as inputs to the MLPNN model. We conducted five different experiments to evaluate the performance of the proposed model in the classifications of different mixtures of healthy segments, epileptic seizure free segments and epileptic seizure segments. We showed that the proposed model resulted in satisfactory classification accuracy rates.

Epileptic seizure detection using probability distribution based on equal frequency discretization.

In this study, we offered a new feature extraction approach called probability distribution based on equal frequency discretization (EFD) to be used in the detection of epileptic seizure from electroencephalogram (EEG) signals. Here, after EEG signals were discretized by using EFD method, the probability densities of the signals were computed according to the number of data points in each interval. Two different probability density functions were defined by means of the polynomial curve fitting for the subjects without epileptic seizure and the subjects with epileptic seizure, and then when using the mean square error criterion for these two functions, the success of epileptic seizure detection was 96.72%. In addition, when the probability densities of EEG segments were used as the inputs of a multilayer perceptron neural network (MLPNN) model, the success of epileptic seizure detection was 99.23%. This results show that non-linear classifiers can easily detect the epileptic seizures from EEG signals by means of probability distribution based on EFD.

Epilepsy diagnosis using probability density functions of EEG signals

In this paper, the equal frequency discretization (EFD) based probability density approach was proposed to be used in the diagnosis of epilepsy from electroencephalogram (EEG) signals. For this aim, EEG signals were decomposed by using the discrete wavelet discretization (DWT) method into subbands, the coefficients in each subband were discretized to several intervals by EFD method, and the probability density of each subband of each EEG segment was computed according to the number of coefficients in discrete intervals. Then, two probability density functions were defined by means of the curve fitting over the probability densities of the sets of both healthy subjects and epilepsy patients. EEG signals were classified by applying the mean square error (MSE) criterion to these functions. The result of the classification was evaluated by using the ROC analysis, which indicated 82.50% success in the diagnosis of epilepsy. As a result, the EFD based probability density approach may be considered as an alternative way to diagnose epilepsy disease on EEG signals.

Gravitational Fuzzy Clustering

Data clustering is an important part of cluster analysis. Numerous clustering algorithms based on various theories have been developed, and new algorithms continue to appear in the literature. In this paper, supposing that each cluster center is a gravity center and each data point has a constant mass, Newtons law of gravity is transformed from m/d2to 1/d2. According to adapted the law, we have proposed novel method called Gravitational Fuzzy clustering. The three main contributions of new algorithm can be summarized as: 1) it becomes more sophisticated technique by taking advantages of K-means, fuzzy C-means and subtractive clustering methods, 2) it removes the dependence on initial condition by taking account of the gravitation effect, 3) it improves the cluster centers by means of the gravity center of clusters. We illustrate the advantage of the resulting of gravitational approach with several examples.

Discretization approach to EEG signal classification using Multilayer Perceptron Neural Network model

Electroencephalogram (EEG) recording systems have been frequently used as the sources of information in diagnosis of epilepsy by several researchers. In this study, rearranged EEG signals were classified by Multilayer Perceptron Neural Network (MLPNN) model. Used data consists of five groups (A, B, C, D, and E) each containing 100 EEG segments. In this study, center points with equal interval were selected on amplitude axis of each EEG segment. EEG signals were rearranged by way of that each amplitude value was shifted to the center point closest to itself. Equal width discretization (EWD) method was used for rearrangement process. Wavelet coefficients of each segment of EEG signals were computed by using discrete wavelet transform (DWT). The mean, the standard deviation and the entropy of these coefficients was used as the inputs of MLPNN model. The model was protected from the overfitting by cross validation. Two different classification experiments were implemented by the same MLPNN model: 1) the classification of healthy volunteers, epilepsy patients during seizure and epilepsy patients during a seizure-free interval, 2) the classification of epilepsy patients during seizure and seizure-free interval. MLPNN model classified EEG signals with the accuracy of 99.60% in first experiment and 100% in second experiment. It is observed that MLPNN classification of EEG signals after rearrangement in amplitude axis provides better results.

Classifying discrete interval densities of EEG signals by using DWT and SVM

This study concentrates on detection of the epileptic activities in the electroencephalogram (EEG) signals. For this aim, features are extracted from the EEG signals by using first wavelet transform and then the approach of densities based on equal frequency discretization, and these features are classified by using support vector machines. The obtained results are compared with the results of three different studies. The results show that the feature extraction method used improves the classification success rate and SVM obtains the highest classification success rate possible in faster running time.

Gravitational approach to supervised clustering for bi-class datasets

There have been many researches about supervised clustering. The problem of common supervised clustering is to train a clustering algorithm by avoiding overfitting. To solve this problem, we develop a new algorithm based on gravitational cluster centers. The novel method avoids overfitting by taking account of the gradient between the misclassification error and the number of gravity centers. Also, it detects the number of gravity centers and their locations from the dataset. Two dimensional synthetic dataset are used in order to provide several viewpoints into this new method. Also, it is tested by using a benchmark datasets.

Supervised Gravitational Clustering with Bipolar Fuzzification

Data clustering is an important part of cluster analysis. Numerous semi-supervised or supervised clustering algorithms based on various theories have been developed, and new clustering algorithms continue to appear in the literature. The problem of common supervised clustering is to train a clustering algorithm to produce desirable clusters and complete clusters over datasets and learn how to cluster future sets of objects. In this paper, we have proposed an algorithm called Supervised Gravitational Clustering based on bipolar fuzzification. Traditional supervised clustering methods identify class-uniform clusters; but the offered method identifies class-multiform clusterswith high probability densities. For this aim we have proposed two approaches: common effect and maximal effect. The first, common effect approach, calculates total effect of all class-centers over searching point. Also, this approach is basis for mapping of novel method. The second, maximal effect approach, determines class-centers with the strongest effect over searching point.