Validation study demonstrates enhanced physical activity classification using body-size-aware chest sensors in youth, highlighting the benefit of personalized developmental features.
Key Points
Develop and evaluate an anthropometry-aware deep learning framework for 22-class pediatric activity recognition using a chest-worn school-badge inertial measurement unit.
Analyzed 4,666 activity records from 128 participants aged 5–18 years using subject-independent five-fold cross-validation.
Captured eight inertial channels (triaxial acceleration, angular velocity, and magnitudes at 100 Hz) integrated with temporal attention pooling and subject context (age, sex, height, body weight, and body mass index).
AATA-DeepConvLSTM achieved an overall record-level accuracy of 0.9377 and a Macro-F1 of 0.9291, exceeding the baseline DeepConvLSTM performance of 0.9092 accuracy and 0.8992 Macro-F1.
Ablation testing showed temporal attention alone improved Macro-F1 to 0.9112, real anthropometric context further increased it to 0.9291, and height alone recovered 0.9264, whereas static postures remained the weakest recognized classes.