The Study Of Heart Sound Pattern Recognition Based On Gaussian Mixture Model

Heart sound signals are one of the important human physiological signals. Heart sound contains a large number of physiological and pathological information about human body. Auscultation is an important non-invasive method of detection and diagnosis for cardiovascular diseases, but auscultation process is easily influenced by external noise interference, auscultation results are also vulnerable to doctor's subjective judgment influence. In addition, heart sound auscultation is greatly impacted with the proficiency of the operator, auscultation skills take longer time to master, so it is difficult to spread the use of it. If we can use simple device to achieve heart sound auscultation, and realize the automatic recognition and classification of pathological heart sound signals, it will have great clinical significance. This subject is really carried out in accordance with this demand.The paper elaborates the production mechanism of normal heart sounds and pathological heart sounds, and the application of pathological heart sounds in heart diseases diagnosis. The heart sounds pattern recognition system includes heart sound data acquisition part and pattern recognition part. The hardware circuit is made to acquire cardiac sounds. Then the acquired heart sounds are preprocessed and the characteristic coefficients are extracted. Finally Gaussian Mixture Model method is used to train and recognize kinds of heart sounds.The main function of heart sound data acquisition circuit is to achieve the acquisition and transmission of cardiac sound signals, using the PC to display the waveforms and subsequent analysis. The acquisition hardware circuit consists of these parts: preamplifier and band-pass filter circuit, 50Hz trap filter circuit, gain control circuit, A/D converter circuit and serial communication circuit between single chip and PC.The pretreatment of acquired heart sounds include de-noising, pre-emphasis, framing and endpoints detection. The key points of the pretreatment are de-noising and endpoints detection. Wavelet threshold de-noising method is used to de-noise in heart sounds. Through analysis the characters of heart sounds and wavelet functions, we choose appropriate wavelet basis. The results make it clear that coif3 wavelet can reach the best de-noising effect. Endpoints detection is an important point before characteristic coefficients extraction. This paper use double threshold detection method based on short-time energy and short-time zero-crossing ratio, the method is simple and easy to carry out.Linear Prediction Cepstrum Coefficients and Mel Frequency Cepstrum Coefficients, which are the best effective coefficients in voice signal recognition, were extracted in this paper. The comparative test illustrates that Mel Frequency Cepstrum Coefficients can reach better recognition performance than Linear Prediction Cepstrum Coefficients, and shorter train time and test time than the latter.Gaussian Mixture Model is used in this paper to train and test heart sound signals. The paper collects 9 kinds pathological heart sounds, including normal heart sounds, mitral insufficiency heart sounds, mitral stenosis heart sounds, aortic insufficiency heart sounds, aortic stenosis heart sounds, ventricular septal defect heart sounds, arrhythmia heart sounds, pulmonary stenosis heart sounds and splitting of heart sounds. The number of heat sound samples is 60. First train a gaussian mixture model for each kind of heart sound, the model is decided by the characteristic coefficients. Then calculate the likelihood function of the trained gaussian mixture models with characteristic parameters to be recognized, the obtained maximum likelihood function value corresponding to the gaussian mixture model is the result of heart sound recognition categories. Experimental results verifies the feasibility of the proposed method for pathological heart sounds pattern recognition, and the recognition rate can reach to 90.6%, which can provide a strong basis for heart diseases diagnosis and clinical application.