Research On Heart Sound Signal Classification Based On Deep Learning

Cardiovascular diseases pose a huge threat to human health.Due to the limited medical resources,many patients cannot be accurately diagnosed in the early stage of the disease,which delays the disease and causes serious damage to the body.Heart sounds reflect the laws of heart movement,and the health status of the heart can be judged by heart sounds.The use of deep learning to realize automatic classification of heart sound signals is of great significance to the society.It can assist doctors to quickly and accurately diagnose cardiovascular diseases,and greatly relieve the tension of medical resources.Therefore,the research topic of this paper is heart sound signal classification based on deep learning,and the main research work is as follows:Firstly,this paper introduces the research significance of heart sound signal classification,summarizes the research status of heart sound signal classification and the defects of traditional heart sound classification methods,and then performs data preprocessing on heart sound signals to better train deep learning models.Secondly,this paper proposes a single-view deep learning-based heart sound signal classification,and constructs a deep learning model Dsa Net to classify heart sounds.In order to alleviate the data imbalance problem of the heart sound dataset,a two-stage training method of decoupled representation learning and classifier is adopted to train the model.At the same time,in order to take advantage of the diversity of heart sound data and increase the number of training samples,random clipping is proposed to obtain more heart sound data,and random clipping is linked with ensemble learning to improve the test accuracy.In the experimental phase,the proposed Dsa Net achieves competitive classification results compared with 7 popular deep learning models.Finally,in view of the shortcomings of the single-view method,this paper proposes a multi-view deep learning-based heart sound signal classification,and constructs a multi-view deep network.The heart sound signal is encoded into a heart sound image using the Grammer angle domain,and Res2 Net and Dsa Net-LSTM are used to extract the features of the two views,and then the features of the two views are fused into the classifier to obtain the classification result.Experiments show that the proposed multi-view deep network achieves the best classification performance compared with 6 baseline models.