Predicting Independent Z-Score Stunting Through Fundamental Anthropometric Measurements Utilizing Extreme Gradient Boosting (XGBoost)

Abstract

Early detection of stunting generally relies on the WHO standard Z-Score calculation, which requires specific instruments, thus limiting diagnostic efficiency in resource-constrained areas. This study proposes a data-centric Machine Learning approach to precisely predict stunting status using only basic anthropometric data (age, sex, weight, and height) without involving Z-Score derived variables. The study evaluated a large-scale post-pandemic dataset consisting of 40,071 medical records of toddlers from Jeneponto Regency, Indonesia.Given the class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied during preprocessing. Comprehensive benchmarking experiments were conducted using XGBoost, Random Forest, LightGBM, and Logistic Regression algorithms. The empirical evaluation results demonstrated that the XGBoost algorithm produced the best classification performance with an accuracy of 97.88%, outperforming Random Forest (95.77%) and registering a significant margin of advantage of 15.36% compared to the baseline linear model.Feature Importance analysis confirmed that Height (52.41%) and Age (28.15%) were the most fundamental predictors. These findings indicate that tree-ensemble-based architectures are capable of capturing complex non-linear correlations between basic anthropometric parameters. The implications of this study offer a scientific foundation for the development of lightweight, low-cost, and automated stunting diagnosis systems independent of conventional Z-Score calculations at integrated health posts in developing regions