Z-Score-Free Stunting Prediction from Basic Anthropometric Measurements Using Machine Learning

Abstract

Early detection of stunting commonly relies on the WHO height-for-age Z-Score, which requires reference standards, calculation tools, and correct interpretation. These requirements can reduce screening efficiency in resource-limited community health services. This study proposes and validates a Z-Score-free machine learning framework for toddler stunting screening using only basic anthropometric measurements: age, sex, weight, and height. The scientific novelty lies in the deliberate exclusion of Z-Score-derived variables, weight-for-age status, weight-for-height status, and other nutritional-status indicators from the predictor set to prevent data leakage. Four yearly datasets from Jeneponto Regency, Indonesia, were combined, producing 40,071 labeled toddler records after removing one blank row. SMOTE was applied only to training data, and missing numeric values were handled through median imputation inside the training pipeline. Four algorithms were evaluated: XGBoost, Random Forest, LightGBM, and Logistic Regression. Under the valid Z-Score-free pipeline, Random Forest achieved the highest hold-out accuracy and F1-score (95.81% accuracy, 94.86% F1-score), while LightGBM achieved the highest ROC-AUC (99.28%) and recall (95.83%). XGBoost remained highly competitive (95.00% accuracy, 93.90% F1-score, and 99.12% ROC-AUC), but it did not outperform all competing classifiers on this dataset. Stratified 5-fold cross-validation confirmed the same overall pattern, with Random Forest producing the highest mean F1-score. SHAP analysis showed that height and age were the dominant contributors to the XGBoost decision process, aligning with the biological basis of stunting as impaired linear growth relative to age.