Examining physical activity clustering using machine learning revealed a diversity of 24-hour step-counting patterns

Area of Science:

• Wearable sensor technology
• Machine learning applications in health
• Physical activity analysis

Background:

• Physical activity is vital for public health.
• Standardized methods for analyzing temporal physical activity patterns are needed.
• This study addresses the need for novel approaches to understand daily activity rhythms.

Purpose of the Study:

• To develop a procedure for clustering 24-hour physical activity patterns.
• To utilize accelerometer-derived step count data for pattern analysis.
• To apply unsupervised machine learning for categorizing physical activity behaviors.

Main Methods:

• Collected 1 Hz step count data from 42 healthy participants using hip-worn accelerometers.
• Applied kernel k-means algorithm with global alignment kernel to 815 days of data.
• Used spectral clustering on 24-hour step-counting pattern probabilities to identify behavioral clusters.

Main Results:

• Identified six distinct 24-hour step-counting patterns.
• Discovered five daily step-behavioral clusters (e.g., all-day dominant, bi-phasic dominant).
• Found that activity tertile groups (high, moderate, low) comprised different proportions of the identified step-counting patterns.

Conclusions:

• Introduced a novel unsupervised machine learning approach for categorizing daily activity.
• Successfully revealed six distinct step-counting patterns and five daily step-behavioral clusters.
• Demonstrated the reliability of this procedure for clustering physical activity patterns and behaviors, highlighting diversity beyond traditional total activity amount categorization.