Mathematical Optimization and Ensemble Learning for Hypertension Risk Prediction: A Feature-Driven Stacked Framework with Interpretability

Abstract

Hypertension remains a leading contributor to cardiovascular disease and premature mortality, yet early risk identification continues to be challenging in many healthcare settings. Traditional statistical models often fail to capture the complex relationships among predictive features, necessitating more advanced, interpretable solutions. This study proposes a mathematically optimised, interpretable ensemble learning framework for hypertension risk prediction, leveraging stacked generalisation and SHapley Additive exPlanations (SHAP)-based explainability. The model integrates four base classifiers-Support Vector Machine (SVM), K-Nearest Neighbours (KNN), Gaussian Naive Bayes (GNB), and Multi-Layer Perceptron (MLP)-with a LightGBM meta-learner. Optuna was used for hyperparameter tuning, and SHAP provided global and local model transparency. The model was evaluated using accuracy, precision, recall, F1-score, Receiver Operating Characteristic (ROC)-Area Under the Curve (AUC), and Precision-Recall (PR)-AUC. The stacked model outperformed all individual classifiers, achieving an AUC of 0.93 (ROC) and 0.96 (PR), with high classification accuracy and balanced sensitivity (recall = 0.86). SHAP analysis revealed KNN_prob and MLP_prob as the most impactful features. Force plots further demonstrated case-level interpretability. The resulting stacked ensemble model offers high prediction performance and interpretability, making it a potential candidate for screening hypertension. Its clinical validity, scalability, and interpretability make it easier to integrate into real-world healthcare systems, especially for early intervention and resource-constrained settings.