Ensemble machine learning for multi-class anemia severity prediction and risk factor identification among Ethiopian adolescents

Background Anemia remains a major public health concern among adolescent girls in Ethiopia, adversely affecting physical growth, cognitive development, educational attainment, and future reproductive health. Accurate prediction of anemia severity and identification of key risk factors are essential for supporting targeted prevention and intervention strategies. Objective This study aimed to develop and compare tree-based ensemble machine learning models for multi-class prediction of anemia severity among Ethiopian adolescent girls using nationally representative survey data. Methods This study analyzed Ethiopian Demographic and Health Survey (EDHS) data from 2011 and 2016, including 15,952 adolescent females aged 15–19 years. Four tree-based ensemble machine learning algorithms (Random Forest, Extra Trees, CatBoost, and XGBoost) were developed and compared with a baseline Multinomial Logistic Regression model. Following TRIPOD-AI recommendations, data preprocessing included missing value imputation, categorical feature encoding, Recursive Feature Elimination (RFE) for feature selection, and Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. To prevent data leakage, all preprocessing and SMOTE procedures were applied exclusively within the training folds. Models were trained using stratified 10-fold cross-validation on 70% of the data and evaluated on an independent 30% hold-out test set (n = 4786) using accuracy, precision, recall, F1-score, and macro-average ROC-AUC. Model interpretability was assessed using global and local SHapley Additive exPlanations (SHAP). Results Random Forest achieved the highest predictive performance, with an accuracy of 85.60% and a macro-average ROC-AUC of 94.84%, followed by Extra Trees (85.10% accuracy; 94.48% ROC-AUC) and CatBoost (83.80% accuracy; 93.08% ROC-AUC). The baseline Multinomial Logistic Regression model achieved 71.20% accuracy and an ROC-AUC of 81.50%. SHAP analysis consistently identified geographic region, altitude, educational level, household wealth index, and household size as the most influential predictors of anemia severity. Conclusion Tree-based ensemble machine learning models, particularly Random Forest and Extra Trees, substantially outperformed conventional multinomial logistic regression for multi-class prediction of anemia severity among Ethiopian adolescent girls. Their strong predictive performance and SHAP-based interpretability highlight their potential to support evidence-based screening, risk stratification, and targeted public health interventions in resource-limited settings.