Predictive Modeling of Child Stunting in Honduras Using Explainable Machine Learning

Abstract

Chronic child malnutrition is a persistent public health problem in Honduras, with irreversible consequences on cognitive and physical development. The aim of this study was to develop and validate predictive models of chronic malnutrition (stunting) using explainable machine learning, to identify the most influential determinants and estimate their individual contribution to risk. Data from 8,713 children aged 0 to 59 months from ENDESA/MICS 2019 were analyzed. A wealth index was constructed using principal component analysis on 21 household assets and three algorithms (logistic regression, Random Forest and Gradient Boosting) were trained with stratified cross-validation. Interpretability was evaluated with SHAP (SHapley Additive exPlanations) values. The prevalence of stunting was 18.9%. Random Forest presented the best performance (AUC-ROC=0.691; AUC-PR=0.362). SHAP analyses identified wealth index as the primary predictor (mean SHAP=0.489), followed by child age (0.323) and maternal education (0.251). A marked socioeconomic gradient was observed (Q1: 39.2% vs Q5: 7.1%), with amplification of the effect in rural areas. In conclusion, explainable machine learning allows the identification and quantification of key determinants of chronic child malnutrition, supporting interventions focused on economic transfers, improvements in water and sanitation, and women's education, with territorial prioritization in the western corridor of the country.