Modelling Hypertension among Adults in South Africa through SMOTE-Based Balanced Data with Machine Learning Approaches

Abstract

Objectives: The objective of the study was to use machine learning (ML) methodology to create prediction models for hypertension based on nationally representative health and demographic surveillance of South African adults, and to consider class imbalance and improved interpretability. Materials and Methods: This was a cross-sectional analytical study utilizing secondary data from Wave 5 of the National Income Dynamics Study (NIDS) covering a total of 21,181 adult respondents across all nine provinces of South Africa. Various ML algorithms were trained and tested. Accuracy, precision, recall, F1-score, and receiver operating characteristic curve's area under it (ROC AUC) were used for model performance evaluation. Feature importance was also investigated by applying SHapley Additive exPlanations (SHAP) for improved interpretability. Results: In the absence of SMOTE, ensemble models attained moderate accuracy (73–75%) but poor sensitivity for classifying hypertensive cases. With SMOTE, their performance greatly improved, and ensemble models of Gradient Boosting, LightGBM, and CatBoost attained perfect classification (accuracy, precision, recall, and ROC AUC = 1.00). Random Forest provided optimum trade-off between accuracy, stability, and interpretability. SHAP analysis identified age, body mass index (BMI), and waist circumference as having greatest influence, followed by sex and lifestyle factors such as smoking and exercise. Conclusion: Incorporation of SMOTE into ensemble ML algorithms significantly improves hypertension prediction model accuracy and sensitivity for South Africa. The study highlights the value of interpretable, data-intensive methodologies for facilitating early identification, focus-dose interventions, and data-informed public health decisions for resource-constrained environments.