Seven Things to Know About Exercise Classification With Inertial Sensing Wearables

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Wearable Sensors

Background:

• Wearable sensors offer potential for remote physical therapy by tracking exercise compliance.
• Current wearables struggle to accurately detect specific physical therapy exercises.

Purpose of the Study:

• Develop open-source classifiers for remote physical therapy monitoring.
• Analyze how data collection choices impact classifier performance.

Main Methods:

• Trained multi-class classifiers using data from 19 healthy adults performing 37 exercises.
• Utilized 10 inertial measurement units (IMUs) placed on various body parts.
• Investigated effects of sensor density, location, type, sampling frequency, and data size.

Main Results:

• Classified 10 exercise groups with 96% accuracy using 10 IMUs.
• Achieved 89% accuracy with a single pelvis-worn IMU.
• Higher sampling frequencies and more data from the same population did not significantly improve performance.

Conclusions:

• Open-source tools and data can lower barriers for developers of exercise monitoring systems.
• Findings support balancing accuracy with wearable form factor for clinical adoption.
• Increased transparency and trust for clinicians and patients in remote monitoring systems.