Bridging Neuroscience and Machine Learning

: EEG Classification for Parkinson’s Disease and Beyond.

Abstract

Electroencephalography (EEG) is a powerful, non-invasive tool capable of measuring brain activity with a high temporal resolution at relatively low cost. Despite extensive clinical use for sleep classification and epilepsy detection, the use of EEG for early-stage neurodegenerative disease detection has been limited. In this thesis, we explore methods for classification of EEG, primarily with a focus on early-stage Parkinson's Disease and develop a novel interpretable neuroscience-informed pipeline that outperforms existing methods.

We begin by introducing a novel dataset comprising healthy controls, early-stage Parkinson's (PD) participants, REM sleep behaviour disorder (RBD) participants, and individuals with both PD and RBD. RBD is of particular interest as studies suggest that a substantial proportion of individuals with idiopathic RBD eventually develop Parkinsonism or cognitive impairment. On the challenging task of early-stage Parkinson's vs. healthy control classification, we implement multiple time series classification methods as baselines that do not integrate neuroscience knowledge. These methods struggle to learn and do not outperform dummy models, thus demonstrating the difficulty of this classification task.

We hypothesise that integration of neuroscientific priors into the machine learning pipeline would improve classification performance. Therefore, we consider brain structure and function in developing a classification approach. We implement frequency bandpowers, peak frequency statistics and connectivity measures into a novel pipeline for EEG classification. Our pipeline significantly outperforms baselines, achieving classification accuracy of 86.2% (±10.4%) and Macro F1 80.7% (±14.1%) for early-stage Parkinson's. Our pipeline provides the added benefit of interpretability via feature importances, which we demonstrate by example with analysis. We find the lightest sleep stage N1 to be statistically significantly more effective for classification vs. the other data types (Wakeful and sleep stages N2, N3, REM) and find Phase Lag Index (PLI) to be the most discriminative connectivity measure.

Finally, motivated by the widespread success of deep learning across fields such as healthcare, natural language processing, and autonomous driving, we explore state-of-the-art deep learning methods for EEG classification. We compare these methods to our neuroscience-informed pipeline. We find deep learning methods perform poorly on the PD classification task, expecting this is due to the low sample size and the difficulty of detecting early-stage Parkinson's. We then adapt our neuroscience-informed pipeline for general EEG task classification, presenting KnowEEG. We apply KnowEEG to five EEG tasks, including mental workload classification, emotion detection and eyes open/closed classification. The datasets for these tasks have been processed to have short signal lengths (2-10 seconds) with many samples (12,000 - 410,000) and have therefore been used for deep learning model benchmarking in the literature. With KnowEEG, we are able to outperform deep learning methods on these tasks and provide the added benefit of interpretable features. By demonstration on the eyes open/closed task, we show that our pipeline uncovers known neurophysiological phenomena correctly, which illustrates great potential for the future. These results pave the way for improved outcomes in many EEG tasks, from disease classification to mental health applications and the brain computer interfaces of the future.

Date of Award	19 Dec 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Raul Santos-Rodriguez (Supervisor) & Zahraa S Abdallah (Supervisor)