CRAFNet: A CNN–Transformer Hybrid Network with Integrated Attention Mechanisms for Heart Sound Classification

As a non-invasive and low-cost diagnostic approach, phonocardiogram (PCG) is widely acknowledged for its effectiveness in facilitating the detection of heart diseases. To enhance the accuracy and robustness of heart sound classification, a deep fusion model named CRAFNet has been developed, integrating residual structures and multiple attention mechanisms. Based on a CNN-Transformer backbone, ResBlocks have been incorporated along with Coordinate Attention (CA) and the Convolutional Block Attention Module (CBAM) to optimize feature weighting in both channel and spatial dimensions. Experimental results on the 2016 PhysioNet/CinC Challenge dataset indicate that the classification accuracy has been improved from 90.48% to 93.11% by CRAFNet, relative to the baseline. An accuracy of 99.2% was additionally achieved on the Yaseen dataset. The effectiveness of the overall architecture has been further validated through ablation studies, underscoring its potential for application in heart sound classification.