A Dual-Accelerometer System for Classifying Physical Activity in Children and Adults

Area of Science:

• Biomedical Engineering
• Epidemiology
• Machine Learning

Background:

• Accurate 24-hour movement behavior monitoring is crucial for time-use epidemiology.
• Previous accelerometer methods faced challenges with wear compliance and activity discrimination.
• Dual skin-attached accelerometers show promise for continuous wear and detailed activity classification.

Purpose of the Study:

• To validate a dual-accelerometer system for classifying physical and sedentary behaviors in children and adults.
• To assess the accuracy of classifying various activities using a dual-sensor approach.
• To compare the performance of dual-sensor versus single-sensor accelerometer protocols.

Main Methods:

• Seventy-five participants (42 children) wore two accelerometers (thigh and lower back).
• Ten standardized activity trials were conducted under direct observation.
• A random forest machine learning classifier was trained using time- and frequency-domain features from accelerometer data.

Main Results:

• Machine learning models achieved high accuracy in differentiating six activity classes (adults: 99.1%, children: 97.3%).
• Using a single accelerometer significantly reduced accuracy for nonambulatory activities (up to 26.4% decline).
• Features incorporating orientation from both sensors were key predictors for accurate classification.

Conclusions:

• The dual-accelerometer system offers a promising advancement for monitoring 24-hour time-use behaviors.
• This validated method enhances the accuracy of classifying diverse physical activities and sedentary behaviors.
• Future research will focus on validating these findings in free-living conditions.