Anthropometry-aware chest-worn inertial sensing for pediatric activity assessment using a school-badge device

Résumé fourni par la source

Objective assessment of physical activity in children and adolescents supports school health surveillance and the evaluation of behavioral interventions, but wearable recognition models are typically developed on adults, public benchmarks, or sample-level data splits that do not test performance on new users. Baseline models also encode the inertial stream alone and carry no description of the individual, leaving developmental differences in body size and movement execution uncorrected. This study developed the Anthropometry-Aware Temporal Attention DeepConvLSTM (AATA-DeepConvLSTM) model for single-label 22-class activity recognition from a chest-worn school-badge inertial measurement unit (IMU). It combines convolutional-recurrent encoding of eight inertial channels (triaxial acceleration and angular velocity at 100 Hz plus their magnitudes), temporal attention pooling, and representation-level fusion of age, sex, height, body weight, and body mass index as subject-level context. After quality control, 4,666 activity records from 128 participants aged 5-18 years were analyzed under subject-independent five-fold cross-validation, with record-level macro-averaged F1 (Macro-F1) as the primary metric. AATA-DeepConvLSTM reached a record-level accuracy of 0.9377 and a Macro-F1 of 0.9291, against 0.9092 and 0.8992 for the strongest baseline, DeepConvLSTM. In ablation, temporal attention alone raised Macro-F1 from 0.8992 to 0.9112 and real anthropometric context raised it to 0.9291, whereas shuffling the anthropometric values across participants returned it to 0.9117; height alone recovered 0.9264. Static postures remained the weakest classes. These results support anthropometry-aware wearable-signal modeling as a feasible engineering approach for pediatric activity assessment, while naturalistic validation and stable calibration protocols remain necessary before routine school-based deployment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.