Ensemble Learning for Pediatric Stunting Detection: A Comparative Study of XGBoost, Random Forest, and LightGBM with Oversampling Techniques

Abstract

Stunting, driven by chronic childhood malnutrition, remains a critical global public health concern. Early detection is persistently challenged by class imbalance in pediatric health datasets and the absence of systematic comparisons between oversampling strategies and ensemble classifiers. This study develops and evaluates an ensemble learning pipeline for stunting detection, benchmarking XGBoost, Random Forest, and LightGBM across five oversampling configurations — Original, SMOTE, ADASYN, Borderline-SMOTE, and SMOTE-ENN — using 10,000 pediatric health records from posyandu activities in Bangka Belitung Province, Indonesia. Seven anthropometric and demographic features were utilized, with stratified 80:20 train-test splitting and five-fold cross-validation. XGBoost with original imbalanced data achieved the highest Recall (0.9573) and a competitive F1-Score (0.9158), while LightGBM with SMOTE delivered the strongest balanced performance (F1-Score: 0.9160, ROC-AUC: 0.8431). SMOTE-ENN consistently underperformed across all classifiers. To our knowledge, this is the first study to simultaneously compare five oversampling strategies across three ensemble models within a unified framework, offering a foundation for high-sensitivity stunting surveillance in resource-constrained healthcare settings.