Ensemble multi-objective hyperparameter optimization for the classification of imbalanced heart disease data

• Developed a novel ensemble method combining NSGA-II and AdaBoost. • Applied multi-objective NSGA-II to optimize multiple classification metrics. • Minimized standard deviation to improve robustness on imbalanced datasets. • Demonstrated effectiveness on MIMIC-III, MIMIC-IV-ED, and HF datasets. Heart diseases remain a leading cause of mortality worldwide, highlighting the urgent need for reliable predictive tools for early diagnosis and risk stratification. In machine learning, selecting optimal hyperparameters is essential for enhancing model performance, especially in contexts of class imbalance and heterogeneous medical data. Metaheuristic algorithms are frequently employed for hyperparameter optimization in both single and multi-objective settings, with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) being one of the most widely utilized. However, most multi-objective-based methods overlook balanced performance across multiple evaluation metrics and often fail to exploit the potential of Pareto fronts fully in hyperparameter optimization. In this study, we introduce the Ensemble Multi-Objective Optimization (EMOO) framework, utilizing NSGA-II with an additional novel objective: minimizing the standard deviation across selected metrics while maximizing them simultaneously. The resulting Pareto-optimal solutions are integrated using Adaptive-Boosting to produce a high-performing ensemble model. EMOO is evaluated on three heart disease datasets–MIMIC-III, MIMIC-IV-ED, and Heart Failure Clinical Records–consistently outperforming baseline methods, SMOTE, and an NSGA-II-based optimization framework. The proposed method achieved superior predictive performance while decreasing variability across metrics. Furthermore, statistical analysis, including Friedman and Nemenyi tests, confirmed significant results. Our findings highlight the value of explicitly incorporating metric-balance and Pareto-front exploitation into multi-objective optimization, offering a promising path toward more stable and accurate classifiers, specifically in healthcare with imbalanced data. EMOO is algorithm-agnostic, allowing base learners, ensemble strategies, hyperparameters, search spaces, and even application domains to be freely adapted beyond healthcare to other imbalanced classification tasks.