Time-to-Fall Prediction in Parkinson’s Disease Using Wearable Sensor Data and Machine Learning

Background Falls are a frequent and serious complication in individuals with Parkinson’s disease (PD), often leading to hospitalization, fractures, and diminished quality of life. While prior studies have focused on classifying fallers and non-fallers, limited research has addressed when a fall is likely to occur. Moreover, few models have integrated dynamic gait metrics obtained from wearable sensors. Objective This study aimed to develop and evaluate a machine learning–based survival model using wearable-derived gait and physiological features to predict time to first fall in patients with PD. Methods We conducted a prospective observational study including 300 PD patients who were monitored over 12 months, among whom 160 (53.3%) experienced at least one fall. Gait parameters such as stride length variability, rhythm irregularity, swing asymmetry, and acceleration variability, along with heart rate variability (HRV), were continuously collected via wearable sensors. Time-to-fall data were recorded, and both Cox proportional hazards and DeepSurv neural network models were constructed. Predictive performance was assessed using concordance index (C-index), integrated Brier score, and time-dependent area under the curve (AUC). Results Early falls were significantly associated with gait instability features (e.g., increased stride variability and rhythm irregularity), lower HRV, diminished balance, and weaker social support. The DeepSurv model outperformed the Cox model, achieving a C-index of 0.73 versus 0.66, with a lower Brier score (0.17 vs. 0.22) and a higher mean time-dependent AUC (0.75 vs. 0.68). Conclusion Integrating dynamic data from wearable devices with survival models enables personalized prediction of fall timing in PD patients. This approach may facilitate timely intervention, particularly in home and community settings, to reduce fall-related morbidity.