Software-engineered Heart Disease Dynamically Weighted Fusion Ensemble Diagnosis: Clinical Decision Support

Background and Motivation: Cardiovascular diseases (CVDs) remain a leading global cause of mortality, with approximately 17.9 million deaths annually. Early and accurate diagnosis is critical to improving patient outcomes and reducing healthcare costs. Traditional diagnostic methods are limited by subjectivity and inability to capture complex physiological patterns, motivating the development of intelligent clinical decision support systems. Proposed Method: This paper presents a dynamically weighted soft-voting fusion ensemble integrating the Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LogReg) classifiers. Unlike fixed-weight ensembles, classifier contributions are adaptively determined by cross-validated F1-scores. Feature dimensionality is reduced using an Analysis of Variance (ANOVA) F-statistic pipeline that retains the eight most clinically significant biomarkers. Experimental Setup and Dataset: Experiments are conducted on the UCI Cleveland Heart Disease dataset, comprising 303 samples (165 healthy, 138 heart disease cases) with 13 clinical features. A stratified 80/20 train-test split was applied, with model optimization performed using Repeated Stratified K-Fold Cross-Validation (5-fold, 2 repeats) to ensure robust and unbiased performance estimation. Key Results: The proposed ensemble achieves 88.33% accuracy, 100% precision, 75% recall, 85.71% F1-score, and 0.93 ROC-AUC on the held-out test set. Compared to the Decision Tree baseline (50% ROC-AUC, 73.33% accuracy), the proposed model achieves a 15-point improvement in accuracy and a 0.43-point gain in ROC-AUC, demonstrating substantial clinical effectiveness and reproducibility.