Epilepsy Seizure Prediction from EEG Signal Using Machine Learning Techniques

Abstract

Automatic seizure prediction is an important task to help epilepsy patients and epilepsy specialists. In addition, measuring electrical activity in different brain parts is an important step before any prediction. The best tool for recording electrical activity is electroencephalography (EEG), which uses electrodes placed on the head. This paper examines the performance of the convolutional neural network (CNN) architectures and support vector machine (SVM) method for predicting epileptic seizure activity using rich information recorded in the signal of EEG segments. The proposed approach is based on 22 features extracted from different EEG segments to produce a representative dataset. SVM classification models and two CNN architectures are proposed to predict ongoing seizures and different states of epilepsy patients. Two CNN architectures are presented: the first is trained with a dataset of features extracted from the EEG signal, and the second is trained with a dataset of Scalogram images from the EEG signal, whose purpose is to predict the imminence of an epileptic seizure in patients. A dataset of 6 patients is used to predict all states of epilepsy patients. Both CNN architectures and binary SVM classifiers achieve a classification rate above 98%.

Keywords

• epilepsy seizure
• eeg
• prediction
• convolutional neural network
• svm