Classification And Recognition Of Heart Sound Signals In Congenital Heart Disease Based On Convolutional Neural Network

Heart sound is an important physiological signal of the human body and is the basis for cardiac auscultation.CHD(congenital heart disease)is a defective disease that is predisposed to the baby at birth.At present,screening for CHD relies mainly on experienced doctors to go to the countryside for auscultation.This method is costly and inefficient,it is difficult to establish a long-term mechanism.By means of digital signal processing,the signal can be analyzed to achieve the purpose of classification and recognition of heart sound signals.The traditional heart sound classification algorithm has poor universality,complexity and low accuracy,which is not good for real-time decision making in the future.This paper analyzes and classifies the heart sounds of CHD and proposes a CHD classification algorithm based on CNN(Convolutional neural network).The algorithm is mainly composed of two parts,heart sound preprocessing method for deep learning and heart sound classification based on CNN.Heart sound preprocessing model for deep learning.The traditional heart sound preprocessing algorithm has a complicated process and the processed heart sound signal has only a time dimension,which cannot meet the deep learning sample processing requirements.This paper proposes a deep-learning heart sound preprocessing algorithm that transforms heart sound signals from one-dimensional vectors to two-dimensional matrices.The heart sound preprocessing model is to retain more effective information and combines the frequency domain characteristics of the heart sound signal to extract the Mel cepstral coefficients into a"sample mapf".The experiment verifies the validity of the preprocessing model and effectively meets the input requirements of the deep learning network.CHD heart sound classification algorithm based on CNN.In order to determine the best classification model,this paper uses 3000 heart sound samples are used to train RNN(Recurrent Neural Network),LSTM(Long Short-Term Memory network),Bi-LSTM,(Bi-directional Long Short-Term),and CNN in different network layers.The results show that the RNN and the Bi-LSTM are prone to over-fitting.The LSTM network loss value of 0.294 and the accuracy of 0.888.The CNN loss value of 0.27 and the accuracy of 0.885.Experiments show that CNN have greater potential than the other three networks.This paper combines the sample size of 14×14×3,34x34x3 and 73x73x3 to optimizes the CNN with 3000 heart sound samples.The experiment shows that the optimal sample size is 34x34x3.In this case,the verification set has an accuracy of 0.908 and a loss value of 0.25.The test set showed an accuracy of 0.91,sensitivity of 0.85,and specificity of 0.97.The algorithm has high accuracy and large sample set,which guarantees the universality of the algorithm.This research is expected to be applied to machine-assisted auscultation.