Epileptic Seizure Detection Using a Small Memory Footprint Convolutional Neural Network

Automatically detecting epileptic seizures has been explored under the light of machine learning, with promising results in the past few years. Epilepsy affects about 51 million people worldwide, while 30% of them do not have control over their symptoms, which causes unpredictable seizures. This means that approximately 15.3 million people will continue to experience seizure episodes even after undergoing treatment (e.g. drugs, surgery). Many previous works have focused on the development of models for patient-specific tasks. This study will explore the problem of seizure detection under the cross-patient approach. We present a technique for generating and preprocessing data from EEG signals that can be used as inputs to a Convolutional Neural Network. The architecture and performance of the network are presented, with results comparable to the state-of-the-art, achieving around 95% average accuracy. The network presented has five convolutional layers, making it a good candidate to be embedded in a wearable device. The method is explained in detail as well as the training hyperparameters.