Towards a Portable Model to Discriminate Activity Clusters from Accelerometer Data.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Data Science

Background:

• Classifying physical activity from accelerometer data is crucial for health monitoring.
• Few existing methods are validated on multiple independent datasets.
• Generalizability of models to diverse populations and settings remains a challenge.

Purpose of the Study:

• To evaluate unsupervised machine learning for developing a reusable and generalizable physical activity classification model.
• To assess the model's performance on independent laboratory and free-living datasets.
• To determine the viability of unsupervised approaches for analyzing accelerometer data.

Main Methods:

• A k-means clustering model was developed using two labeled adult laboratory datasets.
• The unsupervised clustering model was applied to three independent labeled datasets (two laboratory, one free-living).
• Clusters were collapsed into four activity categories: sedentary, standing/mixed/slow ambulatory, brisk ambulatory, and running.

Main Results:

• The model achieved high accuracy (89-96%) for activity categories on development data.
• Consistency of activity types within clusters remained above 70% for sedentary and 85% for ambulatory/running on independent datasets.
• Acceleration features within clusters were consistent across samples, indicating robustness.

Conclusions:

• Unsupervised machine learning offers a potentially useful approach for analyzing accelerometer data.
• The developed clustering model demonstrates reasonable generalizability to diverse datasets.
• This method shows promise for analyzing free-living physical activity data.