Advancing Cardiac Disease Detection Using Feature Extraction, Feature Selection, and Ensemble Learning Approaches

Approximately 26 million individuals globally struggle with cardiac disease, and the incidence is increasing by 2% each year. To reduce the healthcare burden, the researchers propose various CAD models. Feature extraction and Feature selection are essential in reducing the model complexity and memory requirement. In the proposed research, we investigate the performance of different feature extraction and selection methods using two heart sound datasets. The features are extracted using MFCC and DWT methods from heart sounds. Four feature selection methods (Fisher’s Score, mRMR, ReliefF, and Gini Index) are analyzed and ranked using the D-CRITIC TOPSIS technique. The two best models based on feature selection are utilized in the weighted average ensemble. The weights in ensemble learning are optimized using the Dwarf Mongoose optimization algorithm. The feature fusion model combining DWT and MFCC with mRMR for feature selection achieved the highest performance on the PhysioNet dataset, with an accuracy of 82.70%, an F1-Score of 0.8369, and an AUC-ROC score of 0.9092. The best accuracy, F1-Score, and AUC-ROC score on the PASCAL CHSC dataset are 79.64%, 0.7826, and 0.8116, respectively. The study compared four feature selection methods. The mRMR-based model achieved the highest TOPSIS score and ranked first in the performance table. The findings demonstrate that the mRMR feature selection performed better than other feature selection for both feature extraction methods evaluated in this study. The ensemble model using mRMR and ReliefF outperformed all base models and achieved the highest performance metrics. This study highlighted the enhanced detection of cardiac disorders through the combined effectiveness of feature extraction, feature selection, classification models, and ensemble strategies.