Research On Heart Sound Classification Of CHD Based On Deep Learning

Congenital heart disease（CHD）is a problem of structural defects and dysfunction of the heart that appear at birth.The incidence rate is still on the rise in recent years.Yunnan is a high-incidence area of congenital heart disease.For patients with congenital heart defects,their heart blood circulation is in a pathological state,which will affect heart function and work efficiency.With the development of medicine,the key to reducing the mortality rate of congenital heart disease lies in timely diagnosis and effective treatment.However,some remote areas of Yunnan often miss the best treatment time due to limited medical conditions and economic capacity,coupled with weak medical awareness.This article aims at the method of auxiliary diagnosis for the initial diagnosis of congenital heart disease,and studies the use of artificial intelligence to enable people to discover the condition in time without spending a lot of money and material resources.The main research content of the thesis is to analyze and study normal heart sounds and congenital heart disease heart sounds,construct a heart sound signal classification model for congenital heart disease based on deep convolutional neural networks,and classify and recognize the collected heart sound signals to assist in diagnosis.The main contents of the paper is as follows.1.How the heart sound signal is denoised.Aiming at the non-pathological noise doped in the process of heart sound acquisition,spectral subtraction,Butterworth filtering,and wavelet threshold function are used to denoise the heart sound signal.The effect of noise reduction is compared through experiments.The results show that the wavelet threshold function is more advantageous for heart sound denoising.2.How the cardiac cycles are extracted from heart sound signals.Using the medical connection between heart sound and electrocardiogram,a hidden semi-Markov model is built to segment the heart sound signal,accurately locate the four states of heart sound,including S1,systole,S2,and diastole,and extract the value of each heart sound signal several cardiac cycles.3.How heart sound features are extracted.The feature extraction method of heart sound signal based on HMGFCC is proposed,and the windowing part of heart sound is improved to effectively prevent the feature loss caused by spectrum leakage.Aiming at the shortcoming that a single feature cannot better characterize signal features,the advantages of MFCC and GFCC feature extraction are combined to extract the mixed feature HMGFCC,and the principal component analysis algorithm is used for feature selection.4.How the heart sound classification and recognition are made.This paper studies the classification performance of congenital heart disease heart sounds on deep learning models from the perspectives of 1-D and 2-D.First,the single-cycle heart sounds obtained after preprocessing are used as the input of the CNN-LSTM model for training and testing;then the extracted heart sound features are classified in the form of two-dimensional feature maps as the input of the deep learning neural network.The network models used include: improved Alex Net,VGG16,Res Net18,improved Inception network.Studies have found that Inception-v4 has the best performance in heart sound classification.