Research On The Heart Sound Signal Based On Adaptive Time-frequency Analysis

Heart sound signals are one of the most important physiological signals, containing a large number of physiological and pathological information for various parts of the heart, such as atria, ventricle, large blood vessels, cardiovascular. Heart sound signal analysis and recognition is an indispensable means of understanding the state of the heart and blood vessels. In this paper, we carried out an in-depth study in heart sound signal analysis and recognition. The content covers the analysis of heart sound signal, heart sound signal eigenvector extraction, meanwhile, the classification between normal heart sound signals and the four kinds of heart murmurs signal, such as atrial fibrillation(AF), aortic regurgitation(AR), aortic stenosis(AS) and mitral regurgitation(MR). This work includes the following five aspects:a) Heart sound signal acquisition and preprocessing: In this paper, we use the developed stethoscope with a recorder function, to achieve the acquisition of heart sound signal. By analyzing the noise of heart sound signal, wavelet de-noise was selected as a filtering method for heart sound signal. According to the experimental analysis, we choose Garrote threshold function combined with multi-level threshold as the approach for heart sound data pre-processing program.b) Time-frequency analysis of heart sound signals: In this paper, using five kinds of time-frequency analysis method respectively, in order to get the time-frequency spectrum characteristics of heart sound signals. The results show that: Different methods of time-frequency analysis are related to the characteristics of heart sound signals. That is, make a comprehensive consideration between the need for a small cross-term interference and the high time-frequency resolution. In view of this, an adaptive cone kernel time-frequency analysis method was proposed. Experiments show that the distribution better reflect the time-frequency structure of heart sound signals. It's performance is superior to the general CKD and CWD ,SPEC distribution. At last, the method of adaptive cone kernel time-frequency was selected for analyzing the heart sound signals.c) Eigenvector extraction of heart sound signals: According to the analyzed time-frequency results of the standard heart sound data of 3M Littmann?Stethoscopes Database, extracting the 8 groups of features, with the help of Fihser dimensionality reduction approach, the 2-D eigenvector was extracted, which is easy to classify heart sound signals.d) Classification of heart sound signals: According to the composition of heart sound signals scattergram, research on the Support Vector Machine Kernel function, the selection method of multi-classification support vector machine, at the same time, based on the purpose and credibility of classification as the criteria for judging the kernel function parameters and the relaxation variables of the optimization, a support vector machine model was established for classification of heart sound signals. Select 60 heart sound data from 80 heart sound data for each type of group in 2-D eigenvector data with the standard database NM,AF,AR,AS,MR as a support vector machine learning samples, the remaining 20 data were set for test of each type .The classification accuracy (Ar) of each type was 100%. At the same time, the collected clinical data with the above-mentioned same type of 24 cardiac cycles were tested. The classification accuracy of NR,AF,MR was 100% and that of AR,AS was 95.83%, respectively.e) The software system of heart sound signal analysis and classification: In this paper, with the MATLAB visualible function, the software platform for the heart sound signal analysis and recognition was established, which can read the heart sound signal, realize the pretreatment for heart sound signals, draw the three-dimensional and two-dimensional contour map for time-frequency analysis, energy characteristics. simultaneously, using the dynamic-link between MATLAB and EXCEL, the data of heart sound signal analysis were storaged and statistical functions were realized. Finally, through the analysis of heart sound signal 2-D eigenvector, automatic recognition for heart sound signals was achieved.In this paper, the main characteristics of the study was reflected on the extraction method of heart sound signal eigenvector and the establishment of multi-classification support vector machine model. Generally speaking, in this paper, heart sound signals were studied both in the theory and practice. The heart sound eigenvector was mainly extracted by the adaptive cone kernel time-frequency analysis. According to the composition of heart sound signal scattergram, the support vector machine model of heart sound signal classification was established. And normal heart sound signals and the four kinds of heart murmurs were researched in order to obtain a more satisfactory classification results. However, the types of heart murmurs are still insufficient for the classification, therefore, the future work is to focus on collecting more types of heart murmurs in order to further improve the classification accuracy.