Fall Risk Assessment in Stroke Survivors: A Machine Learning Model Using Detailed Motion Data from Common Clinical Tests and Motor-Cognitive Dual-Tasking.

Background: Falls are common and dangerous for stroke survivors. Current fall risk assessment methods rely on subjective scales. Objective sensor-based methods could improve prediction accuracy.

Objective: Develop machine learning models using inertial sensors to objectively classify fall risk in stroke survivors. Determine optimal sensor configurations and clinical test protocols.

Methods: 21 stroke survivors performed balance, Timed Up and Go, 10 Meter Walk, and Sit-to-Stand tests with and without dual-tasking. A total of 8 motion sensors captured lower limb and trunk kinematics, and 92 spatiotemporal gait and clinical features were extracted. Supervised models-Support Vector Machine, Logistic Regression, and Random Forest-were implemented to classify high vs. low fall risk. Sensor setups and test combinations were evaluated.

Results: The Random Forest model achieved 91% accuracy using dual-task balance sway and Timed Up and Go walk time features. Single thorax sensor models performed similarly to multi-sensor models. Balance and Timed Up and Go best-predicted fall risk.

Conclusion: Machine learning models using minimal inertial sensors during clinical assessments can accurately quantify fall risk in stroke survivors. Single thorax sensor setups are effective. Findings demonstrate a feasible objective fall screening approach to assist rehabilitation.