Deep Learning-Based Stride Segmentation With Wearable Sensors: Effects of Data Quantity, Sensor Location, and Task.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Digital Health

Background:

• Accurate stride segmentation is crucial for digital gait assessment using wearable sensors.
• Systematic evaluations of deep learning models for stride segmentation across diverse mobility tasks are limited.

Purpose of the Study:

• To develop and assess Temporal Convolutional Network (TCN) models for stride segmentation.
• To evaluate TCN model performance based on training data quantity, sensor location, and movement complexity.

Main Methods:

• Utilized data from 121 older adults (with and without Parkinson's disease).
• Developed and tested TCN models for stride segmentation.
• Evaluated performance using F1 scores across walking, turning, and stationary/transitional movements.

Main Results:

• Lower limb sensors achieved >95% F1 scores for walking with 5-10 training participants.
• Foot-mounted sensors maintained high performance (99.3% walking, 96.7% turning).
• Wrist sensors showed significant performance degradation, especially during complex movements (e.g., 72.3% turning).

Conclusions:

• Stride segmentation model performance is highly dependent on sensor location and movement complexity.
• Foot-mounted sensors offer robust gait analysis across various tasks.
• Performance during simple walking does not generalize to complex daily movements, necessitating tailored testing.