An Explainable Stacking Ensemble Model for Predicting Childhood Malnutrition among Under-Five Children in Nigeria
Childhood malnutrition remains a major public health challenge in Nigeria, contributing to child morbidity, mortality, impaired cognitive development, and poor long-term socioeconomic outcomes. Although machine learning has shown promise for malnutrition prediction, many models operate as black boxes, limiting their interpretability and adoption in clinical and public health decision-making. This study developed an explainable stacking ensemble model for classifying and predicting childhood malnutrition among children Under-five in Nigeria, using the 2023–24 Nigeria Demographic and Health Survey (NDHS). Data from 9,374 children with valid anthropometric measurements were analysed, with malnutrition status determined using WHO Child Growth Standards based on height-for-age, weight-for-age, and weight-for-height z-scores. Following preprocessing and a multi-method, multi-phenotype feature selection process combining Recursive Feature Elimination, Mutual Information, and Random Forest Feature Importance, twenty-two socioeconomic, demographic, geographic, maternal, and child-health variables were retained for modelling. A stacking ensemble comprising Random Forest, XGBoost, LightGBM, CatBoost, and Support Vector Machine as base learners, with Logistic Regression as meta-learner, was developed using mother-level stratified group cross-validation to prevent clustering-related data leakage. Within-fold SMOTE addressed class imbalance, and Youden's J statistic guided threshold optimisation. Model interpretability was achieved using SHAP, providing both global and individual-level explanations. The stacking ensemble outperformed all individual base learners, achieving a ROC-AUC of 0.750, F1-score of 0.662, recall of 72.6%, precision of 61.0%, and accuracy of 69.9% at the optimal threshold of 0.40. SHAP analysis identified household wealth index, maternal education, child age, and geographic location as the most influential predictors of malnutrition risk. These findings demonstrate that integrating ensemble learning with explainable artificial intelligence provides an accurate, transparent, and practical decision-support framework for early identification of at-risk children, supporting evidence-based nutrition interventions and public health policy in Nigeria.