Abstract: Cardiac auscultation is a basic method for the initial diagnosis and screening of congenital heart disease (CHD). The analysis of heart sound signals is conducive to the realization of machine-assisted auscultation. In this paper, the original heart sound signal was denoised and reconstructed by wavelet transform. The Shannon-envelope of the heart sound signal was extracted and the cardiac cycles were segmented. The feature parameters of MFCC and BFCC with a frame length of 2048 were extracted from S1 as the starting point, and a set of 32-dimensional feature was obtained. It was used as input to the BP neural network for learning, training, and testing of the network. 60 heart sounds of CHD and 60 normal heart sounds were selected randomly for training. Also the heart sounds of 40 CHD and 40 normal heart sounds were selected randomly to test after training. The experimental results show that the sensitivity is 85.11%, and the specificity is 88.89% of BP network by using BFCC features. The network sensitivity of using the MFCC is 86.96%, and the specificity is 90.91%. The BP network by using MFCC feature parameter has a more significant recognition ratio than the one using BFCC feature parameter as the input.Abstract: Cardiac auscultation is a basic method for the initial diagnosis and screening of congenital heart disease (CHD). The analysis of heart sound signals is conducive to the realization of machine-assisted auscultation. In this paper, the original heart sound signal was denoised and reconstructed by wavelet transform. The Shannon-envelope of the heart ...Show More
