Machine learning classification of early-stage Parkinson’s disease using sit-to-walk biomechanical features

Early-stage Parkinson’s disease (PD) impairs motor control during complex tasks requiring coordinated postural adjustment and locomotion. The sit-to-walk (STW) task integrates standing and gait initiation (GI), placing balance demands on people with PD; however, its potential for early PD detection remains underexplored. We aimed to identify STW-related biomechanical biomarkers of early-stage PD and evaluate the performance of a machine-learning-based classification model. We enrolled 106 participants (63 with early-stage PD and 43 age-matched healthy controls). Three-dimensional motion capture, force plates, and surface electromyography assessed participants’ STW task performed at a self-selected speed. We extracted 200 kinematic, kinetic, and neuromuscular variables across three task phases, with Phase 2 (P2) corresponding to the GI phase encompassing the first stepping cycle, during which dynamic balance control is challenged. Weighted feature importance and stepwise binary logistic regression identified three variables: mean center of mass (COM) speed during the entire task; anteroposterior center of pressure-COM displacement during P2; and forward thoracic range of motion during P2, indicating trunk flexion associated with postural adjustment during GI. A random forest classifier incorporating these variables achieved 84.9% accuracy. These biomarkers may be associated with compensatory movement strategies related to postural stability and support objective early screening of PD.