An Algorithm to Classify Real-World Ambulatory Status From a Wearable Device Using Multimodal and Demographically...

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Digital Health

Background:

• Wearable sensor technology enables real-world physical activity monitoring for health insights.
• Accurate measurement of physical activity patterns is crucial for understanding health status.

Purpose of the Study:

• To develop and validate an algorithm for classifying ambulatory status (walking vs. non-walking).
• To evaluate the algorithm's analytical validity and generalizability across demographic groups.
• To assess the algorithm's performance using data from wrist-worn biometric monitoring technology.

Main Methods:

• Algorithm trained and tested on data from two distinct studies with varied ground-truth labeling methods.
• Utilized both high-resolution reference device data (n=75) and participant-reported labels (n=1691).
• A neural network was trained on a combined dataset (16.7 million 10-second epochs) and evaluated on held-out test sets.

Main Results:

• The algorithm demonstrated high accuracy in classifying ambulatory status (AUC 0.938).
• Performance was consistent across demographic subgroups, indicating transdemographic generalizability.
• Daily aggregate metrics showed a mean absolute percentage error of 18%.

Conclusions:

• The developed algorithm accurately classifies walking activity using wrist-worn devices in real-world settings.
• The algorithm exhibits generalizability across diverse demographic subgroups.
• This validated tool can quantify walking activity, providing valuable health status insights for researchers.