The Classification And Application On Heart Sound And Lung Sound Signal

In the field of body data science analysis, cardiopulmonary signals is one of the important information that we care about,which is important physiological information in human body, Heart sounds and lung sounds can be used as important reference for the diagnosis of cardiovascular disease. In the clinical diagnosis, the diagnosis of cardiovascular disease mainly depend on auscultation by doctors, which rely heavily on the experience of doctors. With the development of digital signal processing technology and the application of wearable devices, heart sounds and lung sounds data collection and storage will be more and more convenient.The analysis of cardiopulmonary audio signals data will evaluate health effectively, provide reference for clinical diagnosis, wisdom medical, and improve telemedicine.Most of the research on cardiopulmonary audio signal was heart sounds or lung sounds. Separate research on heart sounds or lung sounds might ignore the identification of some diseases that both heart lesions and lung lesions. The comprehensive research of cardiopulmonary audio signals in this paper, put forward based on the muti-modal support vector machine(SVM) classification model of cardiopulmonary diagnosis. Including acquisition and preprocessing of cardiopulmonary signals data, feature extraction, feature fusion, the construction and application of the cardiopulmonary audio signal diagnosis by muti-modal support vector machine(SVM) model.In this paper, the main work is the following:(1) The collection, preprocessing and standardization of cardiopulmonary audio data. Construct the cardiopulmonary audio database. Self-built database contains 8 kinds of cardiopulmonary audio signal:106 cases of normal heart sounds, 44 cases of aortic stenosis, 63 cases of ventricular septal defect, 65 cases of mitral valve stenosis, 82 cases of mitral stenosis, 69 cases of normal lung sounds, 162 cases of wet rale and 99 cases of wheezing sounds.(2) According to the characteristics of the heart sound and lung sound audio, extracting the audio features in time domain, frequency domain and wavelet decomposition on the three kinds of feature set, including linear predictive cepstral coding(Linear Predictive Cepstral Coding,LPCC) parameters and LPCC’s first-order differential parameters, mel frequency cepstral coefficients(Mel Frequency Cepstral Coefficents,MFCC) parameters and wavelet decomposition coefficient.(3) The feature parameters of cardiopulmonary audio signal fusion algorithm is put forward, to get the feature fusion set in the layer of decision-making and the layer of feature. Experiments show that the fusion of the feature parameters of feature layer classification have higher accuracy than single feature set that LPCC parameters, MFCC parameters and the parameters of the wavelet decomposition coefficient about 72%, 39%, 34% respectively, the recognition rate of normal heart sounds, aortic stenosis, ventricular septal defect, mitral stenosis, mitral insufficiency, wheezing sound are up to 100%. The features fution parameters of decision-making layer classification have higher accuracy result than single feature set that LPCC parameters, MFCC parameters, parameters of the wavelet decomposition coefficient about 62%, 29%, 25% respectively, the recognition rate of mitral stenosis and wheezing sound reached 100%, the recognition rate of normal heart sounds, mitral insufficiency, wet rale reached 92.6%, 95.2%, 97.5%.(4) multi-modal support vector machine(SVM) classification model is put forward, compare of the classification result on the feature of the decision-making level fusion and feature level fusion. Experiments show that multiple modal feature fusion has more advantage than feature set on single-modal state, It have higher recognition rate for each class of cardiopulmonary audio signals, and the model have higher accuracy than the single feature set mode also. The accuracy of the features that fusion on feature layer can reach 89.77%,and the accuracy of the feature that fusion on decision-making layer can reach 80.11%.