Research On Heart Sound Classification And Recognition Algorithm Based On Hybrid Neural Network

Heart sound signal contains abundant physiological and pathological information,reflects the health status of the heart in real time.It is an important evidence for doctors to diagnose cardiovascular diseases.At present,cardiac auscultation is still the main method for the diagnosis of cardiovascular diseases,which is easy to be misjudged due to subjective factors.Combined with the excellent characteristics of the current deep learning algorithm,if we can develop an efficient heart sound classification algorithm to classify and recognize the heart sound signal,it has important reference value for the clinical research of cardiovascular diseases.In view of the above problems,based on the characteristics of heart sound signal,in this paper,two kinds of heart sound classification algorithms based on different hybrid neural networks are proposed to solve the problem of heart sound signal classification and recognition.The main work of this paper is as follows:Firstly,the heart sound denoising method based on improved spectral subtraction is used to denoise the heart sound,and the optimized Mel frequency cepstrum coefficient(MFCC)is taken as the characteristic parameter.Based on the characteristics of heart sound in time domain and frequency domain,a heart sound classification and recognition method based on SDAE and LSTM is proposed.Firstly,the stack denoising autoencoder is used to reduce the feature dimension of heart sound data features,and then the long short-term memory neural network is used to classify and recognize heart sounds.In this paper,we use the open data set for research,set up a number of comparative experiments,and compare the classification results with the relevant literature.The experimental results show that sdae-lstm model performs better than other methods in three classification tasks of normal heart sound,extrastole heart sound and murmur heart sound,and achieves better classification and recognition effect.Then,in order to further explore the influence of feature parameters on the classification and recognition results,that is,on the basis of the original Mel cepstrum coefficient feature,the linear prediction coefficient(LPC)feature is fused,and the cnn-lstm classification model based on hybrid neural network is proposed and constructed.The deep learning model CNN is used as feature extractor,and LSTM neural network is used to classify the heart sounds after feature extraction.On the open data set,the heart sound recognition effect of several classifiers is compared.Compared with other heart sound classifiers,the results show that cnn-lstm hybrid neural network has better recognition effect,and the correct recognition rate of heart sound signal reaches 87.6%.The research results of this paper show that the classification and recognition effect of heart sounds can be improved by fusing the features of heart sounds and combining neural networks,which is of great significance to promote the development of medical treatment and intelligent auxiliary diagnosis.