Segmenting accelerometer data from daily life with unsupervised machine learning.

Area of Science:

• Biomedical Engineering
• Data Science
• Public Health

Background:

• Accelerometers are crucial for measuring physical activity in health research.
• The traditional cut-points method for analyzing accelerometer data is resource-intensive and limits exploration of raw data metrics.
• There is a need for more efficient and flexible approaches to physical activity data analysis.

Purpose of the Study:

• To develop and present a data-driven approach for segmenting and clustering accelerometer data using unsupervised machine learning.
• To overcome the limitations of the resource-expensive calibration studies required by the cut-points approach.
• To enable easier exploration of information from various raw data metrics in physical activity research.

Main Methods:

• Utilized data from 514-year-old participants in the Millennium cohort study wearing wrist accelerometers (GENEActiv).
• Employed a Hidden Semi-Markov Model (HSMM) to segment and cluster data, considering acceleration and angles (x, y, z).
• Compared the HSMM approach with the traditional cut-points method.

Main Results:

• The HSMM approach, using acceleration and angles, identified eight principal components explaining 95% of variance, compared to four with the cut-points approach.
• Distributions of acceleration in HSMM states showed similar groupings to cut-points categories.
• HSMM revealed greater variety in angle distributions compared to the cut-points method.

Conclusions:

• The unsupervised classification approach learns human behavior directly from observed data, eliminating the need for costly calibration.
• This method can integrate multiple data metrics and provides a higher-dimensional description of physical behavior.
• The identified behavioral states are interpretable through observation distributions and duration.