Application of machine learning for hypertension risk prediction across 14 low and lower middle income countries using WHO STEPS data

Hypertension is a major global health challenge, particularly in low- and lower-middle-income countries (LLMICs), where early detection and preventive strategies are often limited. Leveraging machine learning (ML) techniques on population-level data offer a scalable approach to risk prediction and targeted interventions. We used nationally representative WHO STEPS survey data from 14 LLMICs. After preprocessing, and missing data handling, we trained different machine learning models ranging from traditional algorithms (Logistic Regression, SVM) to ensemble methods (Random Forest, XGBoost, Gradient Boosting). To address class imbalance, we applied Synthetic Minority Over-sampling Technique combined with Edited Nearest Neighbors (SMOTEENN). Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). The dataset was split into 80– 20 sets for train test validation and we also compare it with a cross-validation result. On the imbalanced dataset, tree-based ensembles (Random Forest, XGBoost) demonstrated the highest accuracy (0.734–0.74) and AUC (0.75–0.743), indicating robustness. However, these models exhibited low recall (0.261–0.376), reflecting a bias toward the majority class. After balancing with SMOTE, XGBoost and Random Forest improved in accuracy (73.34% and 72.17%) and precision (57.23% and 53.62%), though recall remained low (0.388–0.431). Gradient Boosting showed the most balanced overall performance (AUC: 0.735, F1: 0.520). Naive Bayes underperformed across all metrics, likely due to its assumptions being violated. Cross-validation results demonstrated that XGBoost achieved the best overall predictive performance (AUC = 75.67%, accuracy = 74.73%), while SVM and neural network models showed a more balanced trade-off between sensitivity and precision. SHAP analysis revealed that age group was the most influential predictor of hypertension, followed by sex and educational status, highlighting the dominant role of demographic and socioeconomic factors in the model. Machine learning models can effectively predict hypertension risk in LLMICs, but performance varies notably by algorithm and class distribution. Ensemble models offer strong generalizability but may underperform in recall-sensitive contexts. SVM and Logistic Regression are preferable when sensitivity is prioritized. Our findings highlight the importance of model selection based on use-case requirements and support the integration of resampling and threshold tuning for optimal performance.