A Heart Sound Classification Method Using Cross-contrast Neural Network

Heart sound is produced by the movement of various tissues(valve,myocardium,blood)in the heartbeat cycle,and it contains information from a large number of organs(atrium,ventricle,large blood vessels,valves).We can take effective treatment in the early stages of vascular disease by monitoring the heart sound.Considering the relative shortage of medical resources and our vast territory,the study of automatic auscultation of heart sounds is of great significance to our country's primary health care work.This study combines deep learning and information-based similarity measurement theory(IBS)to propose a new network structure,cross-contrast neural network(CCNN).The network contains two parts.The first part extracts features through a deep network,and the second part uses statistical theory to measure the similarity between feature vectors for classification.This study improves the similarity measurement method in the original IBS theory,and proposes the ModIBS theory,which enables CCNN takes advantange of of deep learning's powerful feature mining ability to classify the signals generated by the dynamic structure based on the statistical and physical assumptions.CCNN mainly has the following main features:1.Using the cross-contrast input mode,on the one hand,extend the medical small data set,and on the other hand,the contrast information other than the signal content information is introduced.2.Using statistical metrics to introduce prior knowledge into the training process of neural networks,so that the network is more suitable for the training of medical small data sets under the support of statistical principles.3.The structure is flexible and easy to adjust,and the methods of feature extraction and distance measurement are alternative.This new network structure makes it combine the advantages of deep network and statistical learning methods.The feature extraction process is simple and contains rich information,which overcomes the difficulty of deep learning in medical applications(training data is small,inter-class differences are not obvious,etc.).Compared with traditional manual design feature methods,the introduction of deep learning in CCNN can simplify the feature extraction process and use machines' computation ability instead of artificial feature design.Compared with the traditional deep learning method,the addition of statistical interpretation makes the whole process easier to understand,and the pair comparison method brings more information,making the network pay more attention to the difference characteristics between different categories.What's more,the combination expand the amount of data,making CCNN suitable for medical diagnosis scenarios with difficult data acquisition and small magnitude,and has strong application potential.This study tested the effect of CCNN on the PhysioNet/Cinc 2016 heart sound dataset with a sensitivity of 0.8346,a specificity of 0.9623,and a mAcc of 0.8985.