Automated seizure detection to support clinical practice

Abstract

Automatic epileptic seizure detection from electroencephalography (EEG) signals has received substantial attention due to its potential to reduce the clinical burden associated with manual EEG review and to support long-term patient monitoring. Despite advances in machine learning and deep learning techniques, challenges remain related to inter-patient variability and the limited interpretability of many high-performing models, which constrains their reliability and acceptance in clinical settings. This dissertation investigates the effectiveness of traditional machine learning, deep learning, and explainable artificial intelligence (xAI) techniques for automatic seizure detection from scalp EEG recordings using the publicly available CHB–MIT Scalp EEG Database. A comprehensive methodology was developed, encompassing signal preprocessing, segmentation, wavelet-based feature extraction using the Discrete Wavelet Transform, and the design of convolutional neural networks (CNNs) operating on both raw EEG segments and Continuous Wavelet Transform (CWT)–based time–frequency representations. Model performance was evaluated primarily using patient-wise cross-validation to provide a clinically realistic assessment of generalisation to unseen subjects. Among traditional classifiers, Support Vector Machines and XGBoost achieved the strongest and most stable performance under patient-wise evaluation, with mean F1-scores of approximately 0.78. Deep learning models demonstrated competitive performance, with a CNN trained on raw EEG signals achieving a comparable mean F1-score of approximately 0.77, while the CWT-based model exhibited lower performance and greater variability across folds. Beyond classification performance, this work placed emphasis on interpretability through xAI techniques. Feature-level analyses showed that traditional machine learning models relied predominantly on high-frequency wavelet components, while Grad-CAM and channel occlusion analyses indicated that CNN predictions were driven by temporally localised patterns distributed across multiple EEG channels. Collectively, these findings support the development of accurate, transparent, and clinically meaningful EEG-based seizure detection systems.