Study on Seizure Detection from the Features of EEG Signals

Determining a seizure is essential in order to support the diagnosis and treatment of an epileptic patient. The objective of this work is to automatically detect the epileptic seizures in a patient from EEG signal. As compared to different biological signals like PET, MEG, MRI and fMRI, the EEG signal is found to be more advantageous. The recorded EEG signal was first preprocessed. Features of the EEG signal were then obtained and finally it was classified into seizure or normal on the basis of the calculated features. The results were compared among the features like Mean, PSE (Power Spectral Entropy), variance and energy and the best performing features were selected. Weighted combinations of these features were obtained in order to confirm a robust feature vector. In this paper, we propose a weighted combination of variance and energies (in two particular frequency bands), resulting in a composite feature. We set a threshold for this composite feature, on the basis of which, a given EEG signal can be classified into a normal or a seizure sequence. The proposed composition of features gives up to 96.5% accuracy.