Explainable AI Identifies High Retrospective Prodromal Symptom Burden in Parkinson’s Disease

Characterizing heterogeneity in Parkinson’s disease (PD) related to prodromal symptom burden may provide insight into early disease mechanisms. Wearable gait analysis offers a scalable approach to capture subtle motor alterations. We developed an interpretable machine learning framework to identify high retrospective prodromal symptom burden (≥3 symptoms) in 275 individuals with PD using single-sensor trunk accelerometry. A three-stage feature selection pipeline identified five key predictors: multiscale entropy (MSE) along three axes, vertical improved harmonic ratio (iHRv), and body weight. A Random Forest classifier achieved ROCAUC = 0.84, PRAUC = 0.86, F1-score = 0.76. Class imbalance was addressed via CTGAN-based augmentation applied within the training set. Explainability analyses consistently identified increased medio-lateral gait complexity (MSEml) and reduced iHRv as dominant contributors. These findings define an interpretable gait phenotype associated with retrospective prodromal burden in PD. While not reflecting true prodromal prediction, this framework provides mechanistic insight into early motor alterations and supports scalable digital phenotyping.