| With the rapid development of economy and changing of human diet, the morbidity and mortality of cardiovascular disease are rising rapidly, which is a serious threat to human health and happy life. Heart sound is a kind of vibration signal which generates by the mechanical activity of heart, and contains abundant diagnostic information of cardiovascular diseases. So heart sound analysis is of great value for non-invasive diagnosis of cardiovascular diseases. As a research hotspot in the field of heart sound analysis, heart sound classification and recognition is intended to identify which type of diseases different abnormal signals belong to according to the feature parameters extracted from different heart sounds. Currently, most of the feature extraction and classification methods are based on linear time-variant or time-invariant models. However, heart murmurs have a nonlinear and non-stationary characteristic, so the linear analysis methods will ignore some inner important information. Given this, a diastolic heart murmurs classification and recognition method is proposed based on the empirical mode decomposition(EMD) in this paper.First of all, on the basis of analyzing the heart sound's and murmur's physiological mechanism and clinical significance, the diastolic heart murmurs are chosen to be as experimental subjects and it can effectively avoid the interference of the physiological murmurs. Aiming at the end effect for the EMD method which is suitable to analyze the nonlinear and non-stationary signal, a new method combining mirror extension and ration extension is proposed to solve the problem. The numerical signals and heart sounds were analyzed, and the results show that the method can effectively reduce the end effect on the EMD decomposition.Secondly, in the aspect of heart sound denoising, the method of wavelet transform is employed. Three important parameters of the wavelet, which are the wavelet basis function, threshold and decomposition scale, are determined though three groups of experiments. In the aspect of heart sound localization, the double threshold method is carried out based on the signal's envelope obtained from the Hilbert transform. In the aspect of feature extraction, the method based on EMD is proposed. On the basis of obtaining the intrinsic mode function(IMF) form EMD decomposition, the criteria of mutual correlation coefficients is used to select the main IMF components. Then Mel frequency cepstrum coefficient(MFCC) and its first-order differential coefficient(?MFCC) and Delta feature are extracted from the main IMF components, respectively. They can form three feature parameters(namely, MFCC, MFCC+?MFCC and MFCC+Delta) which can aliased as E+MFCC, E+M+?MFCC and E+M+Delta through the feature combination. As the EMD decomposition has the problem of mode mixing which can affect the accuracy of the feature extraction, the ensemble empirical mode decomposition(EEMD) is used to take the place of EMD to obtain the IMF components. After selecting main IMF components, three feature vectors(namely MFCC, MFCC+?MFCC and MFCC+Delta) can be obtained, which can aliased as EE+MFCC, EE+M+?MFCC and EE+M+Delta.At last, the hidden Markov model(HMM) which can get reliable model through fewer training samples is employed to be as a classifier. The experiment subjects contain normal heart sound and two kinds of diastolic heart murmurs, namely aortic insufficiency and mitral stenosis. The ratio of the training samples and test samples is 1:2. The HMM model is set up with the feature obtained from the proposed methods to realize heart sound classification and recognition. The results show the recognition performance of the proposed two feature extraction methods is better than the traditional MFCC. At the same time, in order to further verify the proposed end extending method is effective, it was used for the EMD decomposition to extract feature parameters. The experiment results show the recognition rate is higher than the method without end effect processing. |
PDF Full Text Request