Communication and Health in Chronic Childhood Undernutrition: Explainable Ensemble Learning to Identify Dietary Predictors of Infant Feeding Practices and Maternal Supplementation in Ecuador.

内容摘要

Background/Objectives: Chronic undernutrition in the first 1000 days continues to be a significant public health concern in Ecuador. This study examined associations between infant feeding practices, maternal supplementation, socioeconomic factors, and stunting in the first two years of life using explainable machine-learning methods. Methods: The analysis was based on 8613 eligible children aged 0-23 months from the Ecuadorian National Child Malnutrition Survey (Encuesta Nacional sobre Desnutrición Infantil, ENDI) 2023-2024, of whom 8344 had non-missing stunting status and were included in the final modelling sample. Results: Among the children included in the modelling sample, stunting was observed in 18.7%. Maternal iron and folic-acid supplementation were associated with lower stunting prevalence, while breastfeeding and complementary feeding indicators showed weak linear associations but contributed to model predictions in explainability analyses. SHapley Additive exPlanations (SHAP) additionally highlighted geographic and socioeconomic variation, with province, area of residence, and socioeconomic indicators ranking among the features contributing to model prediction. Predictive performance was modest to moderate across models, with area under the receiver operating characteristic curve (AUC) values ranging approximately from 0.62 to 0.66 across the main model specifications. Model performance remained stronger for the majority non-stunted class than for the minority stunted class, highlighting the challenge of detecting stunted children in an imbalanced analytic ENDI sample. Conclusions: Overall, the results suggest that stunting in the analytic sample was represented by multidimensional predictive patterns involving prenatal, nutritional, socioeconomic, and geographic factors, and illustrate the potential value of explainable machine learning for identifying multidimensional risk patterns in child nutrition research.