Research On Heart Sound Detection Method Based On Deep Learning

Cardiovascular disease is the leading cause of death in humans,and the activity of the heart can usually reflect the pathological information of the body.Heart sounds are produced by the sudden closing or turbulence of heart valves,which are important clues for evaluating heart function.At present,a stethoscope is a commonly used clinical diagnostic tool for cardiovascular diseases.When collecting heart sounds,it is susceptible to interference from environmental noise and internal artifacts(such as lung sounds),which affects the effectiveness of doctor auscultation.Since accurate auscultation of the heart requires extensive and long-term training,it is very necessary to use computerassisted heart sound analysis.In this regard,this paper proposes a heart sound segmentation method based on the attention mechanism of CNN and Bi GRU and a heart sound classification method based on heart and lung sound separation and design The intelligent digital auscultation system is developed,and the research content is as follows:(1)The traditional HMM segmentation method,its statistical modeling relies on initialization and insufficient flexibility,while in the deep learning-based method,Bi GRU alone cannot extract the signal local features well.The U-net segmentation method has the problems of down-sampling and loss of information,inability to give more weight to the target that needs attention,difficulty in learning context information,and CE.It is difficult to overcome the state category and various types of heart sound samples.The defect of unbalanced difficulty of state classification.In response to the above problems,this paper proposes a heart sound segmentation method.First,the heart sound envelope signal feature is extracted through CNN,and then the attention mechanism is used to assign different weights to the deep features to achieve expansion and adaptive adjustment of the receptive field,then use Bi GRU to capture context information and FL to solve the problem of sample imbalance.And finally,the sequence of states is obtained by time series modeling.The experimental results show that this method achieves better heart sound segmentation performance than the above-mentioned existing methods.(2)In view of the presence of lung sound interference in the heart sound signal and frequent Mel spectrum I/O operations when reading heart sound features,this paper proposes a heart sound classification method.First,with the help of auxiliary data sets,learn prior knowledge,eliminate the lung sounds in the heart sound signal through the cardiopulmonary sound separation network,and then construct the Mel spectrum equivalent network to avoid repeated Mel spectrum reading,and finally use the heart sound classification network to classify.Joint optimization of the three networks end-to-end.The experimental results show that when the heart sounds are heavily polluted by lung sounds,the use of the heart-lung sound separation network can effectively improve the accuracy of heart sound classification.In addition,this article embeds the above-mentioned heart sound detection algorithm into the intelligent digital auscultation system to provide auxiliary diagnosis.