Predicting childhood anaemia in Ghana with explainable machine learning: A national survey analysis

Introduction Childhood anaemia remains a major public health problem in Ghana, with marked regional and socioeconomic disparities. Conventional regression may not fully capture complex, non-linear relationships among biological, maternal, and household factors. We used supervised machine learning to predict anaemia among children aged 6–59 months using nationally representative survey data. Methods We analysed the 2022 Ghana Demographic and Health Survey, including de facto children aged 6–59 months with valid haemoglobin and complete covariates (weighted N = 3,382). Anaemia was defined as altitude-adjusted haemoglobin <11.0 g/dL. Twenty-one predictors were included. Data were split into training (80%) and testing (20%) sets using stratified sampling. Six models (logistic regression, decision tree, random forest, gradient boosting, support vector machine, and artificial neural network) were tuned via grid search with 10-fold cross-validation. Results The weighted prevalence of childhood anaemia was 48.95% (n = 1,655). Gradient boosting showed the best overall discrimination (AUC = 0.72; F1 = 68.99%; accuracy = 66.27%). Support vector machine and logistic regression achieved the highest sensitivity (recall = 72.73% and 71.74%). Random forest showed overfitting (100% training accuracy; test accuracy = 65.23%). Decision tree and neural network performed poorly (AUC = 0.57 and 0.63). Key predictors across models and SHAP were child age, malaria status, maternal anaemia, region, and household wealth (with feature rankings varying by algorithm). Conclusion Machine learning models achieved moderate predictive performance for childhood anaemia in Ghana. Gradient boosting provided the strongest discrimination, while support vector machine and logistic regression offered higher sensitivity for screening. However, these sensitivities imply that approximately 28–30% of anaemic children may be missed, which should be considered when applying these models in public health screening. Identified determinants support targeted, malaria-integrated nutrition and maternal–child interventions in high-risk groups.