Similarity Gait Networks with XAI for Parkinson's Disease Classification: A Pilot Study

Area of Science:

• Biomedical Engineering
• Neuroscience
• Data Science

Background:

• Parkinson's disease (PD) diagnosis relies on clinical assessments, which struggle to quantify subtle movement changes.
• Objective, quantitative biomarkers are needed for early and accurate PD detection and monitoring.

Purpose of the Study:

• To develop and validate an integrated approach combining graph-based kinematic analysis and explainable machine learning (ML) for identifying digital biomarkers of Parkinsonian motor impairment.
• To quantify alterations in movement dynamics associated with PD using advanced computational methods.

Main Methods:

• Acquired kinematic data from 51 PD patients and 53 healthy controls using Xsens inertial sensors.
• Constructed subject-specific kinematic networks using Jensen-Shannon divergence to model inter-segment similarities.
• Extracted graph-theoretical metrics and applied an ML pipeline (voting feature selection, XGBoost) with nested cross-validation.

Main Results:

• Achieved robust classification performance (AUC = 0.87) in distinguishing PD patients from controls.
• Identified 13 key features using SHAP explainability, highlighting alterations in velocity, inter-segment connectivity, and network centrality.
• Observed increased positional variability, reduced distal limb velocity, and proximal-biased network centrality in PD patients, correlating with clinical severity.

Conclusions:

• The integrated approach effectively identifies quantitative biomarkers for Parkinsonian motor impairment.
• Graph-theoretical kinematic networks and explainable AI offer a promising tool for objective motor assessment in PD.
• These digital biomarkers have the potential to support clinical diagnosis and management of Parkinson's disease.