A machine learning approach to quantifying fall conversion risk in fall-naïve Parkinson's patients

Recurrent falls associated with Parkinson's disease (PD) contribute to diminished quality of life and treatment expense. Current fall prediction models do not adequately predict the transition from non-faller to recurrent faller in people with Parkinson's disease (PwPD). The aim of this project was to develop a machine learning model to identify fall risk to characterize the transition to a recurrent faller.

Baseline clinical motor, biomechanical, cognitive, and quality of life data from 246 PwPD in a clinical trial were combined with subsequent falls reported in diaries over 12 months. Fall-naïve participants (n = 174) were regressed on baseline data to develop a fall conversion prediction model using an XGBoost algorithm. Mean Area Under the Curve (AUC) under repeated cross-validation was the primary performance metric, with a feature importance analysis highlighting contributing variables.

The derived model achieved an across-fold mean cross-validated AUC of 0.63, with mean sensitivity and specificity of 66% and 51%, respectively, and higher sensitivity of 77% for multiple falls, all at a threshold of 0.2. Features contributing to model performance included three biomechanical balance assessments, NeuroQoL lower extremity and MDS-UPDRS part II questionnaires, the Trails Making Test A and the Symbol Digit Matching Test.

The utilization of multifaceted data, biomechanical balance and self-report measures has utility in identifying PwPD at risk of experiencing a first fall and conversion to recurrent faller. Greater precision in identifying the first and subsequent falls has potential to guide interventions and behavior modification to mitigate transition to recurrent faller.

Pragmatic Cyclical Lower Extremity Exercise Trial for Parkinson's Disease (CYCLE-II) TRIAL URL: NCT04000360.