Related Experiment Video
Updated: Dec 12, 2025
06:51
Physical Activity Measurement in Children Accepting Table Tennis Training
Published on: July 27, 2022
2.2K
Machine Learning Models for Classifying Physical Activity in Free-Living Preschool Children
Matthew N Ahmadi1,2, Toby G Pavey2, Stewart G Trost1,2
1Institute of Health and Biomedical Innovation at Queensland Centre for Children's Health Research, Queensland University of Technology, South Brisbane 4101, Australia.
Sensors (Basel, Switzerland)
|August 9, 2020
Summary
Random Forest models using free-living accelerometer data accurately classify preschool children's activities. Optimizing window size and incorporating temporal features significantly improved classification accuracy in real-world settings.
Area of Science:
- * Physical activity recognition
- * Machine learning in human movement analysis
- * Wearable sensor technology
Background:
- * Machine learning (ML) models for activity classification often perform poorly in free-living conditions compared to laboratory settings.
- * Improving ML model accuracy for real-world activity recognition requires adapting training data and model features.
- * Accurate classification of children's physical activity is crucial for health and developmental research.
Purpose of the Study:
- * To evaluate the accuracy of Random Forest (RF) activity classification models for preschool-aged children.
- * To assess the impact of training data from free-living conditions on model performance.
- * To investigate the influence of prediction window size and temporal features on classification accuracy.
Main Methods:
- * Thirty-one preschool children (mean age 4.0 years) wore accelerometers on the hip and wrist during a 20-minute free-play session.
- * Video observation categorized movement behaviors into five activity classes for ground truth.
- * Random Forest models were trained using varying prediction window sizes (1-15s) and with/without temporal features, evaluated via leave-one-subject-out cross-validation.
Main Results:
- * F-scores improved with increasing window size, reaching 86.4% at 15s, with diminishing gains beyond 10s.
- * Inclusion of temporal features enhanced accuracy by an average of 6.2 percentage points, particularly for wrist-based models.
- * Hip and combined hip-wrist models demonstrated superior and comparable accuracy over wrist-only models.
Conclusions:
- * Random Forest models trained on free-living accelerometer data accurately recognize young children's movement behaviors.
- * Optimizing window size and incorporating temporal features are key strategies for enhancing real-world activity classification.
- * Multi-sensor approaches, particularly utilizing hip-worn accelerometers, yield better classification performance than wrist-worn sensors alone.
