Deep Residual Learning for Epileptic Seizure Prediction and Tools to Expedite Biosignal Research

João Miguel Areias Saraiva

Key information

Authors:

João Miguel Areias Saraiva (João Miguel Areias Saraiva)

Supervisors:

Hugo Humberto Plácido da Silva (Hugo Humberto Plácido da Silva); Ana Luísa Nobre Fred (Ana Luísa Nobre Fred)

Published in

11/16/2022

Abstract

Around the world, about 50 million people suffer from epilepsy, a disease characterised by recurrent and unprovoked seizures of abrupt cerebral activity. Epileptic seizures can present a considerable danger to the daily life of patients and those around them. It would constitute a significant improvement in their quality of life if patients could wear some device that would predict their epileptic seizures and raise raise an alarm when a seizure is about to be elicited. In this work, the autonomic nervous system (ANS) deregulation before epileptic seizures - that can be reflected in the patients' electrocardiogram (ECG) - is exploited to predict epileptic seizures. Deep residual learning is proposed to discriminate between ECG segments of seizure eminence - the precital period - from the remaining. ResNet-34 models were train in a patient-specific manner for 11 patients and a cohort median F1-score of 0.728 was attained. A decision algorithm to raise alarms from the segment classifications attained a 0.774 F1-score with a median of 4 false alarms per day. On average, patients that responded to this method had 78.4% of their seizures predicted with a median anticipation of 9.1 minutes. Furthermore, a novel framework for the management and processing of long-term biosignals is proposed to help biosignal engineers to focus more on the research at hands. The LTBio framework achieved a score of 85.75/100 in the system usability scale (SUS) and has proved to be effective in expediting machine learning research with biosignals in the epileptic seizure prediction group.