Epileptic Seizure Detection and Prediction for Patient Support

The identification of epileptic seizure events is acknowledged as one of the most arduous pattern recognition tasks in chronic brain disorders, and has captured considerable interest among researchers. This endeavor has the potential to significantly enhance patients’ quality of life in numerous ways, including accident prevention and mitigating the potential harm associated with epileptic seizures. To enhance seizure detection and prediction efficiency, and provide low complexity pre-trained system, this work presents a trainable hybrid approach to identify seizure events. A shallow autoencoder model has been proposed and implemented to obtain sparse representation of the electroencephalogram (EEG) signal segments followed by the traditional machine learning classifier to categorize the EEG data as either ictal, pre-ictal or inter-ictal. To reveal the ability of EEG channels towards identifying these states, individual channel analysis are presented. Using the CHB-MIT scalp EEG dataset, the proposed method outperforms stae of the art and achieves seizure detection sensitivity of 99.3% and seizure onset prediction sensitivity of 98%. Extremely intensive requirements of computing are minimized by employing a shallow model with fewer parameters to compute and store.