Externally validated machine-learning model using clinical and GAITRite gait data classifies PD fallers vs non-fallers with ~88–89% accuracy and highlights fear of falling, stride/velocity/rotation measures, and autonomic dysfunction as key features.

Although it does not identify molecular therapeutic targets, the work provides a robust, multidomain biomarker for fall-risk stratification and patient selection/outcome measurement in trials aimed at reducing falls or treating balance and autonomic symptoms in PD.

Classification of fallers in Parkinson's disease (PD) is challenging due to the heterogenous motor and non-motor symptoms. We developed a machine learning model integrating clinical and gait data to identify key clinical markers of faller status in PD. Of 468 participants, 396 with complete data were analyzed, with 298 assigned to training from one center and 98 to external validation from the other. Clinical assessments and GAITRite-derived gait metrics were obtained. Fall history classified participants as PD fallers, PD non-fallers, or healthy controls. Features were selected through statistical and importance-based approaches, and seven machine learning algorithms were trained. The Extra Trees classifier utilizing statistics-based feature selection demonstrated the highest performance (accuracy 88% internal, 89% external). Three principal domains consistently emerged: fear of falling (FoF), balance/gait measures (stride length, velocity, 360° rotation), and autonomic dysfunction. These findings support the feasibility of externally validated, multidomain machine learning-based faller classification in PD.