Application of Multi-Domain Feature for Automated Seizure Detection from EEG Signal

Abstract

An automated epilepsy detection method has been proposed by exploiting the multi-domain features with a few learning algorithms. EEG signals are initially preprocessed to remove the redundant data. Then they are divided into 5-second segments, with each segment containing the extraction of multi-domain information from the frequency domain, temporal domain, connectivity, and graph analysis measurements. From the obtained features, most significant features are selected by the Multi Objective Evolutionary (MOE) method. The correlation matrix is obtained through connectivity calculation, and it is converted into binary undirected and weighted graphs through graph theory analysis. For classification, Support Vector Machine (SVM), Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA) has been implemented. Also, the Bayesian optimization (BaO) algorithm has been utilized to optimize SVM parameters. This proposed work is analyzed for EEG signals obtained from the CHB-MIT dataset. The proposed approach resulted in 98.09%, 81.49% and 80.90% accuracy rate for SVM, LDA, and QDA respectively. Conclusively, the SVM classifier outperformed other classifiers in terms of accuracy, Area under the Curve (AUC), sensitivity and specificity with 98.1%, 99.7%, 98.1%, and 98.1% respectively.