Publication

Abstract : Accurately classifying the severity of heart disease into multiple categories is crucial for making well-informed clinical decisions. However, severe class imbalance in datasets such as the Cleveland heart disease dataset poses a significant challenge. To address this, a novel three-stage heuristic-weighted framework is proposed. First, advanced resampling techniques (SMOTE, SMOTE-Tomek, KMeans-SMOTE, and ADASYN) were applied to balance underrepresented severity levels. Second, the multiclass task was decomposed into a tree of four specialized Random Forest classifiers, each trained on data resampled for a specific binary or grouped severity distinction (e.g., "disease vs. no disease", "mild vs. moderate + severe"). Third, their predictions were fused using a transparent, recall-weighted rule-based decision tree that prioritizes detection of severe cases. When evaluated on the Cleveland dataset, the proposed approach achieved an overall accuracy of 77.44%, representing a 1.44% improvement over the current state of the art. Additionally, the method increased recall by 3.11%, significantly enhancing sensitivity to the rarest and most clinically critical classes. By combining systematic oversampling, focused model decomposition, and interpretable fusion logic, this framework delivers a robust, clinically interpretable solution for multiclass heart disease severity classification.