Epileptic seizure classification using AI for multi-modal classification and real-time implementation

Abstract

Epilepsy is a debilitating neurological disorder that significantly impacts the quality of life of affected individuals. Electroencephalogram (EEG) signals are commonly recorded in clinical settings for the study and diagnosis of epilepsy.

This research utilises the Temple University Hospital SeiZure corpus (TUSZ), a dataset consisting of a range of EEG recordings from patients with epilepsy. The dataset includes several seizure types: ABsence SeiZure (ABSZ), Complex Partial SeiZure (CPSZ), MYoclonic SeiZure (MYSZ), Simple Partial SeiZure (SPSZ), Tonic Clonic SeiZure (TCSZ), and Tonic SeiZure (TNSZ).

This thesis focuses on investigating EEG signals for seizure classification. The optimised bagged tree Machine Learning (ML) classifier, using hyperparameter and feature selection, achieved a Weighted F1 score (w-F1) of 0.9828. Comparisons were made with Discriminant Analysis (DA), Naïve Bayes (NB), and Tree classification, achieving w-F1s of 0.9421, 0.7387, and 0.9936, respectively, using two-second windows. Further comparisons using one-second and ictal windows resulted in varying w-F1 scores.

A multi-modal analysis using neurologist reports, a unique component of the TUSZ dataset, was conducted. A Bidirectional Encoder Representations from Transformers (BERT) classifier was used to classify the reports into grouped labels: Focal-all (CPSZ and SPSZ) and Tonic-all (TCSZ and TNSZ), achieving an accuracy of 93.94%. The signals were converted into time-frequency plots and fed into the Inception-ResNet-V2 Neural Network (NN), which decomposed the grouped labels back into their original labels, achieving a label decomposition accuracy of 99.24% for Tonic-all and 99.89% for Focal-all.

Finally, the information from the classical ML approach and the Inception-ResNet-V2 was integrated into a real-time pipeline generated in Simulink, with outputs to C and Field Programmable Gate Arrays (FPGA) for seizure detection and classification. This pipeline achieved four false alarms per day. Although medical applications require only two false alarms per day, the real-time algorithm demonstrated state-of-the-art results.

Date of Award	Jun 2024
Original language	English
Supervisor	Mark Ng (Supervisor) & Jim McLaughlin (Supervisor)