Machine Learning Models for Probability Classification in Spectrographic EEG Seizures Dataset

Manolescu, Vasile Denis and Buckley, N and Secco, Emanuele Lindo (2024) Machine Learning Models for Probability Classification in Spectrographic EEG Seizures Dataset. WSEAS Transactions on Biology and Biomedicine. ISSN 1109-9518 (Accepted for Publication)

Text Denis_v1.pdf - Submitted Version Restricted to Repository staff only Download (1MB)

Abstract

Electroencephalogram (EEG) analysis for seizure detection poses significant challenges due to the complex nature of brain signals. Deep learning models offer the potential to automate this complex process, but their accuracy and reliability remain key challenges. This study aims to enhance EEG seizure probability classification by integrating transfer learning with EfficientNetV2 architectures (i.e., EfficientNetV2S and EfficientNetV2B2), specialised convolutional layers, Long Short-Term Memory (LSTM) units, and fine-tuned attention mechanisms. This approach leverages the strengths of pre-trained models while customising layers to capture the EEG's intricate patterns. The top-performing model achieved 86.45% accuracy with a Kullback-Leibler Divergence loss of 0.95, demonstrating its capacity for accurate and confident seizure predictions. Integrating convolutional and LSTM layers improved performance, effectively capturing spatial-temporal patterns in EEG data. While attention mechanisms offered benefits, hardware limitations were also observed with increasing model complexity. These findings underscore the importance of model customisation against specific dataset characteristics and computational constraints.

Item Type:	Article
Faculty / Department:	Faculty of Human and Digital Sciences > Mathematics and Computer Science
Depositing User:	Emanuele Secco
Date Deposited:	22 Apr 2024 10:10
Last Modified:	22 Apr 2024 10:10
URI:	https://hira.hope.ac.uk/id/eprint/4193

Actions (login required)

View Item