Hip and Wrist Accelerometer Algorithms for Free-Living Behavior Classification.

Area of Science:

• Biomedical Engineering
• Kinesiology
• Wearable Technology

Background:

• Accelerometers are key for objective physical activity (PA) measurement.
• Wrist-worn accelerometers may improve participant adherence compared to hip-worn devices.
• Current methods for analyzing accelerometer data in free-living settings require further validation.

Purpose of the Study:

• To develop and evaluate machine learning (ML) algorithms for classifying PA types.
• To compare the performance of ML algorithms using hip and wrist accelerometer data.
• To assess the validity of ML algorithms against traditional cut-point methods.

Main Methods:

• Forty overweight/obese women wore hip and wrist accelerometers for seven days.
• Wearable cameras provided ground truth activity data.
• A random forest and hidden Markov model classifier was used to categorize activities (sitting, standing, walking/running, vehicle).

Main Results:

• The ML classifier achieved 89.4% (hip) and 84.6% (wrist) balanced accuracy for PA classification.
• ML algorithms accurately estimated daily walking/running time.
• Traditional cut-point methods and laboratory algorithms underestimated walking duration.

Conclusions:

• ML algorithms demonstrate superior performance for PA classification compared to traditional methods.
• While hip-based ML algorithms showed higher accuracy, wrist-based devices may be preferable due to improved participant compliance.
• The developed ML approach offers a promising tool for objective PA assessment in free-living conditions.