Cardiovascular diseases (CVDs) represent the leading cause of worldwide death, requiring precise predictive models for timely intervention. This paper introduces an innovative dualstage stacked ensemble framework that tackles the scalability and reliability constraints of conventional diagnostic methods. The system incorporates four hyperparameter-optimized base learners (Decision Tree, Random Forest, XGBoost, and Logistic Regression) into four unique stacking configurations including diverse meta-learner architectures. Final predictions are produced by soft voting ensemble aggregation to enhance accuracy and reduce clinically significant false negatives. The performance assessment on two heterogeneous widely used datasets exhibits enhanced prediction efficacy. The framework attains 96.22% accuracy with a 3.51% false negative rate on the first dataset and 99.00% accuracy with a 0.893% false negative rate on the second, significantly surpassing current techniques. These findings confirm the effectiveness of multi-level ensemble integration together with thorough hyperparameter adjustment in generating strong, generalizable predictions appropriate for clinical application. The suggested architecture tackles essential healthcare analytics issues while ensuring computational efficiency for practical use.