Wearable Gait Biomarkers and Explainable AI Identify High Retrospective Prodromal Burden in Parkinson’s Disease

Objective wearable markers of Parkinson’s disease (PD) progression and phenotypic heterogeneity remain clinically relevant, particularly in relation to the burden of non-motor features that often precede or accompany the motor diagnosis. In patients with established PD, the retrospective accumulation of prodromal symptoms may identify a clinically meaningful subgroup with broader multisystem involvement. This study developed an interpretable machine-learning framework using wearable inertial gait data to model high retrospective prodromal burden within established PD, defined as the anamnestic presence of at least three prodromal symptoms. A total of 274 individuals with PD performed 30-m walking trials using a single lumbar inertial sensor. Thirty-five biomechanical and clinical variables were extracted, and feature selection identified five key predictors: multiscale entropy ( $$\:MSE$$ ) along three axes, vertical improved harmonic ratio ( \:iHRᵥ ), and body weight. A Random Forest classifier balanced through CTGAN-based training-set augmentation reached a cross-validated ROC AUC of 0.84 ( \:PRAUC = 0.86, $$\:F1-score$$ = 0.76); on the untouched real held-out test set, discrimination was \:ROCAUC = 0.74 and \:PRAUC = 0.71, consistent with an internally developed phenotyping model requiring external validation. Explainability analyses highlighted that increased $$\:MSE$$ and reduced \:iHRᵥ were the strongest contributors to high retrospective prodromal burden, indicating elevated gait complexity and altered spatio-temporal symmetry. These findings delineate an interpretable gait phenotype associated with high retrospective prodromal burden in established PD, supporting wearable gait analysis as a tool for within-PD digital phenotyping.