<![CDATA[Chronic malnutrition known as stunting remains a pressing public health concern in Indonesia, with a national prevalence of 19.8% reported in the 2024 Indonesian Nutrition Status Survey, still well above the 14.2% target set in the 2024–2029 National Medium-Term Development Plan. The Free Nutritious Meal Program (Makan Bergizi Gratis, MBG), launched in January 2025, requires a precise screening mechanism to ensure that nutritional interventions reach the right recipients. The present study is framed as a cross-sectional classification task—rather than a longitudinal predictive model—designed to support point-of-care screening at community health posts. This study proposes an ensemble deep learning framework for early stunting detection that combines three complementary learning paradigms: Random Forest (bagging), XGBoost (gradient boosting), and a Deep Neural Network (DNN). A publicly available Kaggle dataset of 120,999 toddler anthropometric records, in which class labels are likely derived deterministically from the WHO Height-for-Age Z-score (HAZ) formula, was used as the experimental basis. The pipeline included feature engineering grounded in WHO Child Growth Standards, class balancing via SMOTE applied exclusively to the training set, hyperparameter optimization using Optuna, and soft-voting integration of the three base learners. Evaluation was performed on a stratified test set (n=24,200) using accuracy, precision, recall, F1-score, and AUC-ROC, complemented by 10-fold cross-validation and SHAP-based interpretability analysis. The ensemble model achieved 99.84% accuracy, 0.9984 F1-score, and 1.0000 AUC-ROC, exceeding the proposal targets of 88% accuracy and 0.90 AUC-ROC. These near-perfect metrics should be interpreted as evidence that the model has successfully recovered the rule-based labeling structure of the dataset; performance on real-world Posyandu data—where biological variability, measurement error, and unobserved socioeconomic determinants are present—is expected to be lower, and external validation is therefore prioritized as future work. SHAP analysis identified the approximate Height-for-Age Z-score, height, and the height-by-gender interaction as the three most influential features, consistent with WHO anthropometric principles. These findings provide a technical foundation for AI-based screening systems deployable in community health posts and primary care clinics, supporting the effectiveness of the MBG program toward Indonesia Emas 2045.]]>
Copyrights © 2026
