Application Of XGBoost-Based Machine Learning Methods To Predict Stunting

Abstract

Child stunting remains a major public?health challenge across Asia, impairing growth, cognition, and lifelong productivity. Early risk identification is critical, yet conventional screening offers limited predictive power and scalability. This study evaluates machine?learning approaches for stunting prediction using routinely collected infant data, proposing XGBoost and benchmarking it against Logistic Regression and Random Forest. An Asian infant dataset was compiled, label encoding and standardization were applied, class imbalance was addressed with SMOTE, the three models were trained and hyperparameter tuning was performed within a reproducible pipeline. Performance was assessed using Area Under the ROC Curve (AUC) and confusion matrices. XGBoost with SMOTE achieved the highest AUC (0.85), exceeding Random Forest (0.83) and Logistic Regression (0.73). Confusion?matrix analysis indicates that XGBoost separates stunted from non?stunted cases more effectively. Models trained without SMOTE performed worse, underscoring the value of imbalance correction. These findings suggest that ML assisted screening can enable earlier, data?driven risk stratification and targeted interventions. Practical deployment, however, may be constrained by the need for a GPU enabled computer and an IDE based workflow, motivating external validation and implementation refinement.