This research presents a novel hybrid framework for heart disease prediction, integrating advanced data preprocessing with ensemble learning to enhance diagnostic accuracy. The methodology begins with rigorous data cleansing to ensure reliability, followed by Synthetic Minority Over-sampling Technique (SMOTE) to balance class distribution. Outliers are identified and mitigated using the Z-score method, preserving data integrity. A unique Recursive Hybrid Feature Extraction (RHFE) strategy, combining filter and wrapper techniques, optimizes feature selection by reducing multicollinearity and enhancing model efficiency. Key predictive markers include age, chest pain type, maximum heart rate, ST depression induced by exercise, and major vessel count via fluoroscopy. The refined dataset is used to train an CatBoost -based ensemble model, achieving remarkable performance with 94.2% accuracy, 93.5% precision, 94% recall, and an outstanding ROC AUC score of 0.98. These results highlight the model’s robustness and its potential for real-world clinical implementation in early heart disease detection and risk assessment.