An Optimized Heterogeneous Stacking Ensemble with Hyperparameter Optimization for Multi Class Hypertension Risk Classification

Abstract

Hypertension is a major global health concern and a leading risk factor for cardiovascular disease, stroke, kidney failure, and premature mortality. Accurate and interpretable prediction of hypertension risk is essential for supporting early intervention and preventive healthcare. This study proposes an optimized and explainable stacking ensemble framework for multi class hypertension classification by integrating LASSO feature selection, SMOTEENN data balancing, heterogeneous ensemble learning, CatBoost meta learning, GridSearchCV optimization, and SHAP based explainability. The proposed architecture combines five base learners, namely XGBoost, LightGBM, Random Forest, Extra Trees, and Support Vector Machine, whose probability outputs are transformed into meta features and processed by an optimized CatBoost meta learner. Experiments conducted on a dataset containing 3,000 hypertension related records demonstrated superior classification performance, achieving 97.50% accuracy, 97.46% precision, 97.50% recall, and 97.48% F1 score. 10 fold cross validation further confirmed the robustness of the framework with a mean accuracy of 97.50% ± 0.0027. ROC analysis produced AUC values above 0.97 for all classes, indicating excellent discriminative capability. The results demonstrate that the proposed framework provides both high predictive accuracy and strong interpretability, making it a promising solution for intelligent hypertension risk assessment and clinical decision support.