Epileptic seizure prediction using bandpass filtering and convolutional neural network

Abstract

Epilepsy, a chronic neurological disorder, causes seizure- a fast, uncontrollable electrical disturbance in the brain. Seizures that last for a long time might result in memory loss, weariness, photo sensitivity, paralysis, or death. The early diagnosis of seizures may assist reducing the severity of damage and can be utilized to aid in the treatment of epilepsy patients. Predicting seizures before they occur is a challenge that many researchers are working to overcome by monitoring the brain’s activity; but achieving high sensitivity and precise prediction remains a barrier. Our objective is to predict seizure accurately by detecting the pre-ictal state that occurs prior to a seizure. We have used the CHB-MIT Scalp EEG Dataset for our research and implemented the research work using Butterworth Bandpass Filter and simple 2D Convolutional Neural Network to differentiate the pre-ictal and inter-ictal signals. We aim to propose a generalized approach for epileptic seizure prediction rather than patient-specific approach. We have achieved accuracy of 89.5%, sensitivity 89.7%, precision 89.0% and area under the curve (AUC) is 89.5% with our proposed model. In addition, we have addressed several researchers’ seizure prediction models, sketched their core mechanism, predictive effectiveness, and compared them with our work. Our long-term goal is to develop an implantable device to with high accuracy and low errors that may effectively warn patients of oncoming seizures to initiate antiepileptic therapy so that those who are afflicted with the epilepsy can enjoy a healthy and risk-free life.