Deep learning-based classification with improved time resolution for physical activities of children.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Pediatric Health

Background:

• Childhood obesity is a growing global epidemic with long-term health consequences.
• Existing caloric balance methods are often unsuitable for children.
• Accurate monitoring of physical activity is crucial for intervention.

Purpose of the Study:

• To develop and evaluate a novel approach for monitoring children's physical activities.
• To utilize a convolutional neural network (CNN) for real-time activity classification.
• To provide an accurate and applicable tool for childhood obesity prevention efforts.

Main Methods:

• 136 children (8-12 years) wore waist-mounted accelerometers during various activities.
• Accelerometer data was preprocessed into 2.8-second segments.
• A CNN model was trained on approximately 183,600 data samples to classify ten distinct physical activities.

Main Results:

• The CNN achieved 81.2% accuracy classifying ten activities, improving to 91.1% when similar activities were merged.
• The CNN outperformed conventional algorithms like SVM, DT, and kNN.
• Activity merging significantly enhanced performance metrics, including recall, precision, and F1 score.

Conclusions:

• The developed CNN algorithm effectively distinguishes children's physical activities using accelerometer data.
• Merging similar activities improved classification accuracy and mitigated performance degradation.
• The algorithm's simplicity and accuracy make it suitable for real-time applications in childhood obesity management.