A Preliminary Mechanics-Informed Machine Learning Framework for Objective Assessment of Parkinson's Disease and Rehabilitation Outcomes.

Background/

Objectives

Non-invasive methods for evaluating rehabilitation outcomes in Parkinson's disease (PD) remain limited. This preliminary study proposes a mechanics-informed machine learning (ML) framework integrating force-plate data with dimensionality reduction, clustering, and statistical analysis to objectively assess motor control and the effects of a targeted intervention.

Methods

Twelve PD patients were randomly assigned to a PD control group performing standard exercises or an intervention group incorporating additional transverse-plane trunk motion exercises for 10 weeks. Ground reaction forces and center of pressure (COP) signals were recorded pre- and post-intervention using a force plate, alongside data from six healthy individuals as a benchmark. Features related to postural sway and COP dynamics were extracted and refined using Forward Feature Selection. Dimensionality reduction (t-SNE) and unsupervised clustering (K-means) identified group-level patterns. SHAP values and Cohen's d quantified feature importance and effect size. Clustering robustness was assessed with bootstrapping, nested cross-validation, and permutation testing.

Results

K-means clustering revealed clear pre/post-intervention separation in five of six intervention patients, with post-intervention states shifting toward the control cluster. Clustering showed strong performance (Silhouette 0.77-0.79; Calinski-Harabasz 100.8-184.9; Davies-Bouldin 0.29-0.45). The most predictive features (RMS-SML and PL-SAP) showed large effect sizes (Cohen's d = -12.1 and -4.53, respectively) distinguishing PD patients from healthy controls. Traditional statistical tests (e.g., ANOVA) failed to detect within-group changes (p > 0.05), but ML-based methods captured subtle, nonlinear postural adaptations.

Conclusions

This preliminary mechanics-informed ML framework detects PD-related motor deficits and rehabilitation-induced improvements using force-plate data, warranting validation in larger cohorts.