<p>Heart disease prediction remains a critical challenge due to the presence of heterogeneous clinical features and complex nonlinear relationships. Existing machine learning and deep learning models often suffer from limited generalization, static feature handling, and lack of adaptive decision mechanisms. To address these issues, this study proposes a dynamic weighted ensemble framework integrated with feature aggregation for improved heart disease prediction. Initially, relevant clinical features are identified and aggregated using statistical and clustering-based analysis. Multiple base learners are then trained, and their predictions are combined using a loss-driven softmax weighting strategy, where weights are dynamically adjusted based on individual model performance. Furthermore, deep feature extraction is performed using a transformed input representation to capture complex feature interactions, followed by classification using an artificial neural network. The proposed model is evaluated using the UCI heart disease dataset with rigorous tenfold cross-validation. Experimental results demonstrate superior performance, achieving 97.5% accuracy, 98.5% precision, 98.9% recall, and 98.5% F1-score. The results confirm that the proposed framework improves prediction reliability, reduces overfitting, and enhances generalization compared to existing methods.</p>