Fall Risk Assessment in Stroke Survivors: A Machine Learning Model Using Detailed Motion Data from Common Clinical...

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Clinical Biomechanics

Background:

• Falls pose a significant risk for stroke survivors, with current assessment methods often relying on subjective measures.
• Objective, sensor-based approaches are needed to enhance the accuracy of fall risk prediction in this population.
• Developing objective tools can aid in personalized rehabilitation strategies for stroke survivors.

Purpose of the Study:

• To develop machine learning models for objective fall risk classification in stroke survivors using inertial sensors.
• To identify optimal sensor configurations and clinical test protocols for accurate fall risk assessment.
• To provide a data-driven approach to supplement subjective fall risk evaluations.

Main Methods:

• 21 stroke survivors underwent balance, Timed Up and Go, 10 Meter Walk, and Sit-to-Stand tests, with and without dual-tasking.
• Kinematic data were captured using 8 motion sensors, from which 92 spatiotemporal gait and clinical features were extracted.
• Supervised machine learning models (SVM, Logistic Regression, Random Forest) were trained to classify high vs. low fall risk.

Main Results:

• A Random Forest model achieved 91% accuracy in fall risk classification.
• Key predictive features included dual-task balance sway and Timed Up and Go walk time.
• Models utilizing a single thorax sensor demonstrated performance comparable to multi-sensor configurations.

Conclusions:

• Machine learning models with minimal inertial sensors can accurately quantify fall risk in stroke survivors during clinical assessments.
• A single thorax sensor setup is effective, simplifying data collection.
• This objective screening approach shows promise for guiding stroke rehabilitation efforts.