Parkinson's disease (PD) is characterized by alterations in movement dynamics that are difficult to quantify with conventional clinical assessment. This study proposes an integrated approach combining graph-based kinematic analysis with explainable machine learning to identify digital biomarkers of Parkinsonian motor impairment. Kinematic signals were acquired using Xsens inertial sensors from 51 patients with PD and 53 healthy controls. For each participant, subject-specific kinematic networks were constructed by modeling inter-segment similarities through Jensen-Shannon divergence, from which global and local graph-theoretical metrics were extracted. A machine learning pipeline incorporating voting feature selection, and XGBoost classification was evaluated using a nested cross-validation design. The model achieved robust performance (AUC = 0.87), and explainability analyses using SHAP identified a subset of 13 features capturing alterations in velocity, inter-segment connectivity, and network centrality. PD was characterized by increased positional variability, reduced distal limb velocity, and a redistribution of network centrality towards proximal body segments. These features were associated with clinical severity, confirming their physiological relevance. By integrating graph-theoretical modeling, explainable artificial intelligence, and machine learning methodology, this work provides a method of discovering quantitative biomarkers capturing alterations in motor coordination. These findings highlight the potential of ML and kinematic networks to support objective motor assessment in PD.
Bianco et al. (Wed,) studied this question.