Study On Feature Extraction And Classification Recognition Of Congenital Heart Disease Based On AR Model Parameter Spectrum Estimation

Congenital heart disease has brought great impact to human life and health, and its incidence has been at a higher level. According to incomplete statistics, each year about1.5million new patients with congenital heart disease in the world, and nearly sixty percent of patients can not get so timely treated that state of an illness was delayed and cause to die. However, if all patients can be timely treated, most patients can recover. Currently, clinical diagnostic methods has some shortcomings, but the heart sound signals can show the pathological information of the state change of body and heart disease. We can analyze the signal by digital signal processing method and provide reference for clinical diagnosis.Based on MATLAB, we will analyze the signal from three aspects, such as preprocessing, feature extraction and recognition in the paper.Preprocessing include de-noising and segmentation. First the wavelet threshold de-noising algorithm is used to eliminate noise and compare the effect on de-noising effect of different threshold and wavelet base. Then we adopt the positioning method without reference signal to realize the segment of heart sound signals. Firstly, normalized average Shannon energy algorithm is used to get the envelope of signal. Secondly, the envelope combined with their own characteristics to identify the exact location of first heart sound and second heart sound.Feature analysis and extraction includes the signal of time-frequency analysis and the power spectrum estimation. Application of Short-time Fourier transform and Choi-Williams distribution to compare the differences between normal and abnormal heart sound signals. The characteristics of signals are extracted about time, frequency and intensity by power spectrum estimation based on AR nodel.The support vector machine (SVM) method is used to identify diagnosis. Recognition effect is compared by using radial basis function (RBF), polynomial function and Sigmoid function as the kernel function. Through the experiment found that the classifier built by RBF has the highest recognition rate,it can reach76.7%.