Non-Anthropometric Childhood Obesity Risk Screening in Indonesia: An Explainable Machine Learning Approach

Abstract

Obesity in children aged 5–12 years is a pressing public health challenge in Indonesia, where Riskesdas 2018 reported a combined overweight and obesity prevalence of about 20%. This study presents an explainable AI framework as a proof of concept for screening childhood obesity risk from non-anthropometric features, integrating TabNet, Optuna, and a dual-explainability analysis (attention and SHAP) with an exploratory fairness evaluation across SES and region. Using the Indonesia Family Life Survey Wave 5 (IFLS5, 2014/2015) with 8,077 children from 13 provinces, a four-scenario Feature-Set Ablation Study (Scenarios A–D) was designed. Anthropometric features produced a spurious AUC of about 1.0000, consistent with data leakage. In Scenario D (non-anthropometric), no model was superior across all metrics and performance remained limited. Logistic Regression achieved the strongest predictive performance, whereas TabNet was selected as the explainable framework because it natively handles missing survey data and provides intrinsic attention-based interpretability, capabilities that Logistic Regression and Random Forest structurally lack. The analysis identified health status, sex, age, socio-economic quintile, and digital device ownership as the most influential predictors. The framework is positioned for validation on national data rather than as a final screening instrument.