A Comparative Analysis of Machine Learning and Deep Learning Approaches for Phonocardiogram Classification Using Dataset Integration

The worldwide increase in cardiovascular diseases and related deaths necessitates advanced, non-invasive diagnostic methods that allow for timely detection and intervention to mitigate the risk of death. This study analyzes phonocardiogram (PCG) heart sound (HS) classification using Machine Learning (ML) and Deep Learning (DL) techniques. To improve accuracy and data diversity, two well-known datasets—PhysioNet CinC Challenge 2016 and CirCor 2022—were fused after resampling and label harmonization. ML algorithms like XGBoost and LightGBM, and DL algorithms like CNN+BiLSTM Attention and CBAM ResNet were incorporated to train the models leveraging the Mel-frequency ceptral coefficients as key acoustic features. Evaluation was performed using the stratified k-fold cross-validation with bootstrapping utilized multiple orthogonal metrics like sensitivity, specificity, F1- score, ROC-AUC alongside accuracy computed over all folds. The results analysis revealed 95.21% accuracy for the ResNet + CBAM model the best performers followed by CNN + BiLSTM at 94.01%. While ML models provided quicker inference and interpretability, DL models exhibited better generalization as well as greater sensitivity to intricate HS, particularly the quadruple rhythm. The findings reinforce the importance of dataset fusion and evaluation of hybrid models for the creation of clinically usable diagnostic systems. With this hybrid model approach, the proposed framework has advanced responsive capability for early detection of CVD even in constrained healthcare environments.