Advanced machine learning techniques for clinical EEG classification

Zhu, Yixuan

Abstract

In the rapidly evolving field of medical diagnostics, the accurate classification of electroencephalogram (EEG) signals remains a crucial task for the early detection and treatment of neurological disorders. Traditional EEG analysis relies on manual feature extraction, which struggles with the high dimensionality, nonstationarity, and noise inherent in EEG data. Furthermore, labelled clinical EEG datasets are often sparse, limiting the performance of conventional supervised learning. To address these challenges, this thesis investigates advanced machine learning techniques—particularly deep learning, transfer learning, and multimodal integration—for enhancing EEG classification. The central research question is: how can machine learning be used to enhance the accuracy and robustness of normal versus abnormal EEG classification in real-world clinical screening scenarios? Originally motivated by the challenge of cognitive function assessment in multiple sclerosis (MS), the research pivots to a more foundational and clinically relevant task: distinguishing normal from abnormal EEG recordings. This classification task is not disease-specific but serves as a foundational step in EEG screening workflows, enabling early detection of abnormal neural patterns regardless of underlying pathology. As such, it holds particular importance for triage, referral prioritisation, and supporting diagnostic workflows in resource-limited clinical environments.

The first contribution introduces a multi-stage model architecture based on meta-modelling principles. This design is motivated by the observation that traditional windowing methods, although improving sensitivity, often suffer from biased labels and a lack of global context. By optimising window length and overlap strategies, the proposed approach mitigates label noise and captures broader temporal dependencies, achieving 99.0% accuracy on the Temple University Hospital Abnormal EEG Corpus (TUAB) and 86.7% on the more diverse Temple University Hospital EEG Corpus (TUEG), nearing the ceiling set by inter-rater agreement.

The second contribution adapts state-of-the-art audio classification models—Patchout Spectrogram Transformer (PaSST) and Learnable Audio Frontend (LEAF)—to the EEG domain. This direction addresses two critical issues: the performance bottleneck of small-scale models and the data limitations faced by larger architectures. By leveraging pre-training with pseudo-labelled EEG data and transferring inductive biases from audio, PaSST achieves 96.1% accuracy on TUAB, setting a new benchmark for single-stage EEG classification.

The third contribution develops a multimodal model that integrates EEG signals with clinical textual reports. This is motivated by error analysis of existing models, which reveals frequent misclassifications caused by a lack of patient-specific background information. Mimicking the diagnostic reasoning of physicians, the model incorporates raw text embeddings to provide contextual information, leading to a state-of-the-art accuracy of 97.9% on TUAB.

Despite these advancements, key challenges remain. Future work should focus on addressing data scarcity, enhancing model interpretability, improving computational efficiency, and mitigating window labelling biases. The use of self-supervised learning is also proposed to deepen multimodal integration, enabling more nuanced and clinically meaningful interactions between EEG signals and textual data.

• Thesis Type: Thesis
• Deposit Date: Feb 6, 2025
• Publicly Available Date: Jun 9, 2025
• Public URL: https://uwe-repository.worktribe.com/output/13722720
• Award Date: Jun 9, 2025