Application of Explainable Artificial Intelligence for personalized childhood weight management using IoT data

Childhood obesity is a growing global health concern as it is correlated with an increased risk of adult-onset and chronic diseases. Recent advances in digital healthcare technologies have enhanced the efficiency of health data analysis and diagnosis, leading to increased interest in artificial intelligence (AI) applications in childhood obesity research; however, several challenges remain, such as data limitations, class-imbalance issues, and difficulties in model interpretability. This study addresses these challenges through a comprehensive framework that utilizes wearable devices for real-time lifestyle data collection and employs Wasserstein generative adversarial networks (WGANs) to address data imbalance concerns. The proposed framework incorporates an explainable model architecture combining tabular attention network (TabNet) with eXtreme Gradient Boosting (XGBoost), complemented by SHapley Additive exPlanations (SHAP) analysis for enhanced interpretability. The framework was validated using data collected from 362 elementary school students over six months, with an additional external validation set of 82 students. The results demonstrated exceptional performance with 98.0% accuracy on the test dataset and 85.2% accuracy on the external validation data. Therefore, the framework can provide personalized health guidance by identifying and explaining individual factors that contribute to weight change, thereby enabling targeted intervention strategies for childhood obesity prevention.