<p>The increasing availability of patient BioData, including clinical measurements and physiological indicators, offers unprecedented opportunities for developing intelligent, data-driven diagnostic tools. In the context of cardiovascular disease (CVD)—the leading cause of mortality globally—mining such BioData effectively is critical for enabling early detection and supporting complex clinical decision-making. However, traditional predictive models often fail with inherent trade-offs, such as balancing predictive accuracy across imbalanced classes, minimizing feature redundancy, and ensuring model interpretability. To address these limitations, this study introduces a two-stage prediction framework for heart disease. First, a Multi-Objective Genetic Algorithm (MOGA) is employed to perform optimal feature selection by simultaneously maximizing classification accuracy and minimizing redundancy among the 13 clinical features of the UCI Cleveland Heart Disease dataset. Second, the selected features are used to train a deep ensemble of regularized Multi-Layer Perceptrons (MLPs). The ensemble outputs are aggregated using uniform weighting and further refined through AdaBoost-based fusion to enhance robustness. This integrated approach ensures that the model captures clinically meaningful patterns across diverse patient profiles, while also improving interpretability for medical practitioners. Experimental results demonstrate that the proposed framework achieves 96% accuracy, 97% sensitivity, and an AUC-ROC of 0.978, outperforming several baseline machine learning models. The findings confirm the potential of the proposed MOGA–MLP ensemble framework for real-world clinical deployment, offering a reliable, interpretable, and generalizable solution for cardiovascular risk prediction.</p>