Research On Diagnosis Of Electrocardiogram And Phonocardiagram With Deep Learning Algorithm

Cardiovascular diseases,including myocardial infarction,arrhythmia and other diseases,have become the biggest killer of people’s health problems due to their high incidence.Heart beat is a major way to reflect cardiovascular disease.Electrocardiogram(ECG)and phonocardiogram(PCG)are simple and effective tools for measuring the normality of heartbeat,which can provide more assistance for the diagnosis of cardiovascular diseases,and even directly diagnose certain diseases.This thesis mainly studies the automatic diagnosis of cardiovascular disease based on heart beat.The diagnostic framework is based on ECG and PCG,including three major modules: signal preprocessing,feature extraction and model classification,as well as sub-modules such as noise removal and signal segmentation.Algorithmic improvements and innovations have been made in multiple points of the diagnostic process.The ECG and PCG databases used in this article are from CinC Challenge 2017 and CinC Challenge 2016,both of which are public databases from the PhysioNet website.In this thesis,two signal segmentation methods,RR interval segmentation and fixed segmentation,are analyzed.It is pointed out that RR interval segmentation is insufficient for deep learning model,while fixed segmentation will lose the original signal,and a sliding window segmentation method is proposed,which improves the generalization ability of the model.In this thesis,the existing CNN model is improved on the application of ECG and PCG.By analyzing the difference between onedimensional signal and two-dimensional image,a large-scale convolution kernel onedimensional residual network(LKNet)is proposed.In the ECG diagnosis stage,the difficulty level of the four types of label classification in the ECG database was analyzed.Combined with the idea of the external penalty function in the optimization problem,the loss function is modified in a targeted manner.In the PCG diagnosis stage,a Multi-Feature Sets Convolutional Network model(MFS-CNN Boosting)was proposed.Finally,the diagnostic architecture proposed in this thesis has an F1 value of 84.06% on the ECG database and a MAcc value of 91.48% on the PCG database.The state-of-art effect is achieved on both databases.