Activity Recognition for Persons With Stroke Using Mobile Phone Technology: Toward Improved Performance in a Home...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Digital Health

Background:

• Smartphones offer potential for monitoring post-stroke physical activity via built-in sensors.
• Current activity recognition (AR) systems often use data from healthy individuals in lab settings.
• Post-stroke gait impairment and home environment differences necessitate validation of AR for stroke patients.

Purpose of the Study:

• To evaluate AR performance in a home setting for individuals with stroke.
• To compare AR performance using different training data origins: population (healthy vs. stroke) and environment (lab vs. home).

Main Methods:

• Thirty individuals with stroke and 15 healthy subjects performed activities in lab and home settings wearing smartphones.
• A custom app collected sensor data (accelerometer, gyroscope, barometer); subjects self-labeled activities.
• A random forest AR model was trained using healthy or stroke activity data, comparing performance metrics like recall and misclassification.

Main Results:

• AR models trained on stroke data achieved higher average recall (75%) compared to those trained on healthy data (53%).
• Healthy-trained classifiers showed reduced performance with increased gait impairment, misclassifying ambulatory activities.
• AR models trained on in-lab data had lower recall for at-home activities (56%) than for in-lab data (77%).

Conclusions:

• High-quality AR for gait-impaired stroke patients requires training data from this specific population.
• Incorporating at-home activities into training datasets is crucial for effective AR systems in home and community monitoring settings.