Determinants of childhood and adolescent obesity: An explainable AI approach using the ICAD database.

OBJECTIVES

To identify the biological, behavioural, and socio-demographic determinants of childhood and adolescent obesity and evaluate how their relative importance shifts across age strata using an Explainable Artificial Intelligence (XAI) approach.

STUDY

DESIGN

A cross-sectional analysis of a large-scale, multi-country pooled database.

METHODS

Data from 9660 participants (aged 5-17 years) from the International Children's Accelerometry Database (ICAD) were analysed. A CatBoost machine-learning classifier was implemented to predict overweight and obesity. Sparse Principal Component Analysis (SPCA) was utilised to address multicollinearity among features. SHapley Additive exPlanations (SHAP) were employed to quantify the global and local contributions of predictors, providing transparency to the model's findings.

RESULTS

The model achieved robust predictive performance (PR AUC: 0.8003). Systolic blood pressure emerged as the primary predictor of weight status, followed by HDL cholesterol and insulin-related markers. Age was identified as a key determinant and was inversely associated with obesity risk. Additionally, moderate and moderate-to-vigorous physical activity (MPA and MVPA) contributed significantly to the model, demonstrating greater predictive utility for individuals categorised as overweight or obese.

CONCLUSIONS

An XAI framework integrating CatBoost and SPCA successfully identified cardiometabolic biomarkers and physical activity as primary predictors of youth obesity. The analysis uncovered critical discrepancies between WHO standards and ICAD data. These findings advocate for BMI-specific and developmentally tailored public health strategies to optimise paediatric obesity prevention and monitoring.