Epileptic Seizure Classification for Each of Five Classes Using Time-Domain and Wavelet Decomposition Based Features

With the rapid developments in medical technologies and computers, it is getting easier to detect neurological disorders. However, the early detection of Epilepsy which is a common and important neurological disorder is still a crucial and challenging procedure. Like the other disorders, early detection gives some chances for controlling and smoothing of the effects of the disease. Developing early detection methods for epileptic seizures from Electroencephalography (EEG) signals using signal processing and machine learning techniques has been a popular research area in last decades. In this work, we propose to use some time-domain features together with same features extracted for sub-signals of the original signal. For obtaining sub-signals from EEG signal, we used classical discrete-time wavelet decomposition. All EEG signals have been decomposed to five important sub-bands. After extracting features from EEG signals and all sub-signals, we have applied the feature matrices to some machine learning algorithms. Most of the researches in this area distinguish only two or three states of epileptic seizures. In contrary to majority, we preferred to classify all five classes from EEG signals. Classification results of our method show an important success comparing with the other researches. Using statistical, auto-correlation based and wavelet transform based features of EEG signals together to predict and detect epilepsy makes our method robust. In this work, we present the comparison of machine learning techniques for the detection of epilepsy using electroencephalography (EEG) signals. The five states of epilepsy are classified using Naive Bayes (NB), Decision Tree (DT), Random Forest (RF), Multilayer Perceptrons (MLP), k-Nearest Neighbors (k-NN), and Support Vector Machines (SVM) algorithms. Results show that Random Forest, Multilayer Perceptrons, k-Nearest Neighbors (k-NN) with 10-fold cross-validation are the most successful algorithms with 92% accuracy.

I. Introduction

Epilepsy is a crucial neurological disorder which affects more than 50 million people around the world. [1]. Detection of epilepsy is a challenging problem because it is determined by recurrent seizures, and it is not easy to detect its symptoms without monitoring brain activity through EEG. Epileptic seizure which is described as a short event of hyperactivity of neurons can be pre-ictal, inter-ictal, and post-ictal [2–4].