Personalised Consumer Wearable Data for Near-Term Fall Risk Classification in Community-Dwelling Older Adults

Introduction
Current fall risk assessment tools are largely episodic, clinician-led, and not suited to continuous, real-world monitoring. Consumer wearables could offer continuous passive monitoring. Prior studies using wearable data use clinical-grade sensors or lab-based gait tests so limited real-world applicability Consumer wearables (Fitbit, etc.) are widely adopted but underexplored for fall risk classification. 
 

Method
This was an observational cohort study over 6 months with 140 Fitbit users. Nested cross-validation: outer 10-fold × 5 repeats, inner 10-fold for hyperparameter tuning 6 models compared: 4 sequential logistic regression baselines + XGBoost wearable-only + XGBoost full The model classified each 7-day window as fall-associated, if it preceded a reported fall, or non-fall-associated.

Discussion: 

Combined (wearable + clinical) model significantly outperformed all baselines, showing wearable data adds value beyond clinical features alone. Wearable-only performance was comparable to the best clinical-only model, suggesting consumer wearables capture genuine fall-risk signal. Feature contributions across movement, sleep, cardiovascular, and clinical domains support a multidimensional view of fall risk. Within-person baseline modelling may detect changing vulnerability that episodic, point-in-time assessment can miss. Scalable wearable monitoring could help identify emerging risk between clinical contacts and support timely preventive action