Sensor-enabled Activity Class Recognition in Preschoolers: Hip versus Wrist Data.

Area of Science:

• Biomedical Engineering
• Kinesiology
• Wearable Technology

Background:

• Traditional cut-point methods for accelerometer data analysis have limitations.
• Pattern recognition approaches offer a promising alternative for processing accelerometer data.
• Limited research exists on pattern recognition for physical activity in preschoolers.

Purpose of the Study:

• To develop and test activity class recognition algorithms using accelerometer data in preschoolers.
• To compare the performance of algorithms trained on hip, wrist, and combined hip and wrist accelerometer data.
• To evaluate supervised learning algorithms for physical activity classification in young children.

Main Methods:

• Eleven preschoolers (3-6 years) wore accelerometers on the hip and wrist during 12 physical activity trials.
• Activities included sedentary, light games, moderate-to-vigorous games, walking, and running.
• Random Forest and Support Vector Machine classifiers were trained using time and frequency domain features from 15-second windows.

Main Results:

• Support Vector Machine models achieved higher accuracy (up to 0.85) when using combined hip and wrist data.
• Random Forest models also showed improved accuracy (up to 0.82) with combined sensor data.
• Excellent accuracy was observed for sedentary activity, moderate for games and running, and modest for walking.

Conclusions:

• Machine learning algorithms effectively predict physical activity classes from accelerometer data in preschoolers.
• Combined hip and wrist accelerometer data yield superior recognition accuracy compared to single-site placement.
• This approach enhances the validity of accelerometer-based physical activity assessment in young children.