Automatic Epileptic Seizure Detection: Graph F eatures Versus Graph Kernels

Abstract

According to WHO 2019 announcement, around 50 million people are suffering from epilepsy worldwide. As epilepsy causes some seizures in the brain, seizure detection can play an essential role in treating patients. In this paper, we concentrated on different graph-based methods intending to classify seizure and non-seizure states of the brain based on recorded EEG signals. We worked on Temple University Hospital (TUH) dataset which includes both focal and generalized seizures. Our goal was to reach a comprehensive comparison between these methods. Three methods were discussed: graph features, graph kernels, and graph multi-kernels. We considered each EEG channel as a node in the graph model. Also, graph edges were built through functional connectivity between every two nodes' signals. Therefore, we constructed one graph for each second of every patients' recorded EEG. Then, by using constructed graphs, we extracted some features from them, or calculated kernel matrix for each couple of them which reflects the similarity between graphs. In the multi-kernel method, these two approaches gathered together. After comparing the outcomes, we found kernel and multi-kernel methods more effective on this dataset. The best result is attained by multi-kernel method which has an accuracy of 72.1 % and a sensitivity of 71.9%.