Prediction of activity type in preschool children using machine learning techniques

Area of Science:

• Pediatric physical activity monitoring
• Machine learning in child health
• Wearable sensor technology

Background:

• Machine learning accurately predicts activity from accelerometers in adults and adolescents.
• Preschool-aged children's physical activity patterns are crucial for development.
• Developing accurate activity recognition for young children is an unmet need.

Purpose of the Study:

• To develop and validate machine learning models for predicting activity type in preschool children.
• To compare the performance of standard Artificial Neural Networks with Deep Learning Ensemble Networks for this task.

Main Methods:

• Eleven children (3-6 years) performed 12 standardized activities while wearing a hip-mounted accelerometer.
• Activities were classified into five categories: sedentary, light, moderate-to-vigorous, walking, and running.
• Feed-forward Artificial Neural Networks and Deep Learning Ensemble Networks were trained on accelerometer data features.

Main Results:

• The Deep Learning Ensemble Network achieved higher overall accuracy (82.6%) compared to the standard Artificial Neural Network (69.7%).
• The Deep Learning Ensemble Network demonstrated improved recognition accuracy across all activity classes, particularly for light activities (91%) and moderate-to-vigorous activities (79%).

Conclusions:

• Ensemble machine learning, specifically Deep Learning Ensemble Networks, can accurately predict activity types in preschool children using accelerometer data.
• This approach holds promise for objective physical activity assessment in young children.