Machine learning models for predicting childhood anemia in Mozambique: analysis from national survey data

BACKGROUND: Childhood anemia remains a major public health concern in sub-Saharan Africa (SSA), with Mozambique among the most affected countries. Despite the growing use of machine learning (ML) to enhance disease prediction, there is a lack of national-level evidence on its application to childhood anemia in low-resource settings. This study aimed to develop, compare, and interpret ML models to predict anemia among children under five years of age in Mozambique using nationally survey data. METHODS: Data were obtained from the 2022–2023 Mozambique Demographic and Health Survey (MDHS). The study included children under five years of age. Anemia was defined as a hemoglobin concentration below 11.0 g/dL. Five machine learning models were developed and internally validated: Logistic Regression, Random Forest, XGBoost, LightGBM, and CatBoost. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC–ROC) and calibration curve. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) analysis. RESULTS: Among the 9,289 children included in the analysis. CatBoost showed the highest discriminative performance (AUC–ROC = 0.914), with a sensitivity of 87.6% and a specificity of 73.5%. SHAP analysis identified maternal age, child’s age, household wealth, number of children in the household, and lack of vitamin A supplementation as the most influential predictors of anemia. CONCLUSION: Overall, the ensemble boosting models demonstrated the highest discriminatory performance, with CatBoost achieving the best results. These findings suggest that interpretable and low-cost predictive models may support early screening and help guide targeted interventions for childhood anemia; however, their use in practice should be approached with caution. Further work is required, including external validation in independent populations, as well as regional recalibration and fairness assessment, to ensure robustness, generalizability, and equitable application in real-world settings.