Monitoring Physical Behavior in Rehabilitation Using a Machine Learning-Based Algorithm for Thigh-Mounted...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning for Health

Background:

• Physical activity monitoring is crucial for health outcomes.
• Accelerometers are key tools for tracking physical behavior.
• Multi-sensor systems offer detailed insights but face compliance challenges.

Purpose of the Study:

• Develop and validate a machine learning algorithm for classifying daily physical behaviors.
• Utilize a single thigh-mounted accelerometer to simplify data collection.
• Enhance the accuracy of physical behavior recognition in research and clinical settings.

Main Methods:

• Trained a random forest model with 11 behavior classes (sitting, lying, standing, walking, running, cycling, stair climbing, wheelchair ambulation, vehicle driving, transitioning).
• Validated the algorithm using simulated free-living conditions with chest-mounted cameras for ground truth.
• Compared the algorithm's performance against a vector threshold-based method.

Main Results:

• The algorithm achieved an average F-measure of 57% for 11 distinct physical behaviors.
• Performance significantly improved when behaviors were merged into broader categories (sedentary, standing, walking, running, cycling).
• The refined algorithm demonstrated high performance compared to ground truth and existing methods.

Conclusions:

• A single thigh-worn accelerometer can effectively classify specific physical behaviors.
• High classification accuracy was observed for behaviors common in higher-functioning populations.
• Future research should focus on classifying behaviors in individuals with lower functional capacity.