A Lightweight Self-Supervised Contrastive-Inspired Framework for Interpretable Parkinson’s Disease Staging using Wearable Gait Signals

Parkinson's disease (PD) is characterized by progressive neurodegenerative changes resulting in movement impairments, instability during gait, shaking, and problems with posture. Accurate staging of the disease is necessary for timely and efficacious treatment interventions; however, current standardized clinical assessments are often subjective, time-consuming and have considerable inter-rater variability. Although the integration of wearable sensing technologies with AI for automated gait analysis has become possible, these approaches typically require supervised learning, large training sets of labelled data and have limited interpretability in their models.

This study presents a lightweight, self-supervised, contrastive-inspired approach to interpreting PD staging through wearable gait data. The framework incorporates data augmentation and a lightweight multi-layer perceptron (MLP) as an encoder to produce latent representations of augmented gait data and applies classical ML classifiers in combination with explainability via SHAP to a unified architecture. By producing discriminative latent embeddings from augmented gait data that do not require extensive amounts of labelled data, our framework effectively captures comorbidity associated with disease progression through gait data. These latent embeddings are classified using Random Forest (RF) and Support Vector Machine (SVM) classifiers, with RF yielding superior classification performance within the context of our proposed framework.

Results from the experimental study demonstrate RF and SVM classifiers using our framework achieved ~95% accuracy across all three classes while outperforming the baseline ML classifier and being more robust to imbalanced class distribution. The PCA biplots indicate that our framework produced clearly defined embeddings between the three classes, while the confusion matrix indicates that the multi-stage classification results were accurate. Additionally, SHAP-based explainability identified clinically relevant gait features that were related to the progression of the disease and thereby enhanced the transparency of our framework, leading to improved trust in clinical decision-making.

The current framework provides an efficient, interpretable, and reliable system for stage classification of Parkinson’s disease, thus, it can be used as a solution to monitor disease progression in real time and to allow for wearable health monitoring.