Interpretable Machine Learning Models for Predicting Hypertension Risk

Abstract

Hypertension ranks among the most significant contributors to cardiovascular disease, stroke, and premature mortality worldwide. Because the condition typically produces no noticeable symptoms until serious complications have already developed, identifying at-risk individuals early is critical. Traditional statistical approaches often struggle to capture the complex, nonlinear interactions among demographic, clinical, and behavioural variables, and although many machine learning methods outperform these classical models, their opaque, "black-box" nature limits clinician trust and slows adoption in practice. This study develops an interpretable machine learning framework for predicting hypertension risk using Framingham Heart Study records comprising 4,240 participants and 13 attributes. Following data cleaning, imputation, standardisation, and feature selection, four ensemble algorithms — Decision Tree, Random Forest, Gradient Boosting, and XGBoost — were trained and optimised through GridSearchCV on an 80:20 stratified split. XGBoost produced the strongest overall results, achieving a ROC AUC of 0.9582, accuracy of 0.8939, precision of 0.8078, recall of 0.8631, and an F1-score of 0.8346. Explanations generated through SHAP, LIME, and decision-tree visualisation consistently identified systolic and diastolic blood pressure as the leading predictors, a pattern broadly consistent with established clinical diagnostic criteria. Grouping predicted risk scores into Low, Medium, and High categories allowed the framework to correctly flag 86.31% of individuals in the High Risk group. Taken together, the results show that strong predictive performance and model transparency can be achieved simultaneously, providing a practical and reproducible tool for early hypertension risk assessment in clinical practice