Abstract: The changes of heart sounds can represent the signs of heart diseases earlier, and the classification of heart sounds based on deep learning has the function of clinically assisting noninvasive diagnosis of heart diseases. Adhering to the concept of simple feature acquisition and simple deep learning model, a ResNet152 network model integrating CBAM attention mechanism, Focal Loss function and multi-scale features is designed to perform deep learning on the PhysioNet/CinC 2016 heart sound data set. The CBAM attention mechanism, fusion mode in network bottleneck structure, Focal Loss function principle and multi-scale feature acquisition mode are introduced in detail, and five groups of comparative ablation experiments and transverse comparison experiments are designed. Experimental results show that the CBAM attention mechanism, Focal Loss function and multi-scale features improve the classification accuracy of baseline network ResNet152.

Key words: Mel-frequency coefficient characteristic, convolutional block attention module(CBAM), Focal Loss, multi scale