Explainable Multi-Level Modelling for Parkinson’s Disease Severity Assessment Using Wearable Sensor Data

Parkinson's disease (PD) is a progressive neurodegenerative disorder that severely affects the quality of life of older people. Wearable technologies combined with machine learning methods have shown significant progress in the quantitative assessment of PD, but they still face challenges in moving from controlled settings towards free-living environments. Specifically, wearable-derived features are often fragmented and lack clinical interpretability, single-form representations fail to capture complex PD severity patterns, selecting and integrating suitable classifiers for different representations remains challenging, and data scarcity with strong subject-specific effects further limit model performance.

To address these issues, this thesis proposes an explainable multi-level modelling framework for PD severity assessment using short-term motor task data collected from wearable sensors in a low-supervision, semi-controlled setting, as a step towards free-living environments. The framework covers feature, representation, decision, and subject-level modelling. At the feature level, a structured multi-scale and multi-domain feature assessment framework is systematically proposed to extract explainable motor features. At the representation and decision levels, a multi-form fusion strategy integrates statistical features, raw signals, and time-frequency and recurrence-based images using suitable classification models, including Light Gradient Boosting Machine (LightGBM), residual network (ResNet), and convolutional neural network (CNN) models. At the subject level, a structured data augmentation strategy is proposed to leverage latent subject-level representations to support data augmentation. Overall, this work provides a systematic methodological foundation for explainable wearable-based PD severity assessment. With task-specific adaptation and independent validation across datasets, devices, and patient populations, the proposed multi-level framework may also support wearable assessment of other chronic diseases.