Epileptic Feature Extraction via Convolutional Neural Networks

Abstract

Epilepsy, the fourth most common neurological disorder in the world, affects millions of people worldwide. Despite its prevalence the neurophysiological basis of epileptic seizures is poorly understood, complicating the monitoring and treatment of the disorder. Typical seizure detection and classification is performed by a trained epileptologist or neurologist and requires sifting through hours of electroencephalography data. To provide life-saving clinical treatment to epileptic patients and push forward understanding of this disease, methods for automatic detection and prediction of seizures are needed. In this thesis, a convolutional neural network is used as the basis for an automatic seizure detection system that operates on raw EEG data with no pre-processing. Hyperparameter optimization of the proposed system is explored; the system is tested pre- and post-optimization on the CHB-MIT pediatric seizure dataset.