Classification of Heart Sounds Using MFCC and CNN

Abstract

An excellent heart sound classification system can be used as a good method to complete the daily heart sound detection under the condition of low cost and high efficiency, which is convenient to detect problems in the early stage of heart disease, and at the same time can alleviate the problem of medical staff shortage. In this paper, the data set of 3000 heart sound samples was used to accomplish the system. The noise reduction and feature extraction of cardiac sound signals, as well as the construction and training of network models was used to describe in detail. In order to improve the accuracy of the system, contrast the coif5 wavelet and db6 wavelet noise reduction effect, without segmentation of one mind under the condition of the second heart sounds, compared the LPCC, MFCC and the effect of pure heart sound wave input, used for feature extraction has advantages of convolution neural network classification task, and the loss function and optimizer batches for cross training, finally realizes a in 3000 samples with 98% accuracy of heart sound classification system.