Title : Explainable wearable gait biomarkers identify high prodromal burden in Parkinson’s disease
Abstract:
Identifying Parkinson’s disease (PD) during its prodromal phase remains one of the major unmet challenges in neurology. While non-motor symptoms such as hyposmia, REM sleep behavior disorder, depression, and constipation have been incorporated into prodromal research criteria, their individual specificity remains limited. Cumulative prodromal burden may better reflect underlying network vulnerability and early neurodegenerative spread. Parallel to non-motor changes, subtle gait alterations may occur before overt motor disability and may serve as objective digital biomarkers of early disease-related dysfunction.
This study developed an interpretable machine-learning framework using single-sensor wearable gait data to identify high prodromal burden in individuals with established PD, defined as the presence of three or more prodromal symptoms. A total of 275 individuals with idiopathic PD performed standardized 30-meter walking trials using a lumbar inertial measurement unit positioned at L5. Thirty-five biomechanical and clinical features were extracted, including spatiotemporal parameters, harmonic ratios, recurrence quantification metrics, and multiscale entropy along three axes.
A structured three-stage feature selection pipeline identified five stable predictors: medio-lateral, antero-posterior, and vertical multiscale entropy, vertical improved harmonic ratio, and body weight. To address class imbalance, Conditional Tabular Generative Adversarial Networks (CTGAN) were used for synthetic minority augmentation while preserving multivariate structure. A Random Forest classifier achieved robust performance (ROC AUC = 0.84, PR AUC = 0.86, F1-score = 0.76).
Explainability was ensured through SHAP analysis, interaction assessment, calibration analysis, decision-curve evaluation, and surrogate tree modeling. Medio-lateral multiscale entropy (
) emerged as the dominant predictor, with higher entropy values strongly associated with increased probability of high prodromal burden. Reduced vertical harmonic symmetry (
) provided complementary discriminatory information. Importantly, groups did not differ in global motor severity, suggesting that the identified biomechanical signature reflects a specific prodromal-related phenotype rather than generalized disease progression.
These findings support the use of interpretable wearable gait biomarkers to quantify cumulative prodromal burden and advance digital phenotyping strategies for early stratification in Parkinson’s disease.
