A method for heart sound classification using sample augmentation and INDANet

The number of deaths due to cardiovascular diseases has been steadily increasing, with the majority occurring in underdeveloped regions. Doctors in less developed regions often have limited clinical experience to diagnose these diseases by listening heart sounds, whereas artificial intelligence can serve as a valuable tool to assist them in conducting auxiliary diagnoses. However, small samples and strong noise reduce the diagnosis accuracy. Therefore, in this paper, a novel method is put forward for heart sound classification, which integrates sample augmentation and injected noise dual attention networks (INDANet). First of all, the heart sounds are preprocessed through a Butterworth filter to eliminate noise outside the cardiac frequency range. Then, a sample augmentation is utilized to increase the size of sample set. In addition, a suitable dose of Gaussian noise is injected to improve the robustness and generalization of INDANet with channel and spatial attention mechanism. Experiments on two datasets demonstrate that the proposed method achieves superior performance in heart sound classification compared to the other six advanced models. The accuracy in the two datasets achieves as high as 99.85% and 98.07%, respectively.