Research On Classification Of Heart Sound In Small Samples

Heart sound classification can be used to detect abnormal heart sound and find patients with cardiovascular diseases(CVDs).It plays an important role in the diagnosis of CVDs.It is a major research hot-spot in the field of automatic diagnosis of CVDs based on heart sound and has attracted the research interest of many domestic and foreign scholars.The traditional methods of selecting heart sound features by experience to perform classification tasks have some shortcomings,such as difficulty in obtaining effective features and poor classification effect.With the development of artificial intelligence technology,deep learning has become the mainstream method to study the heart sound classification in recent years,and the classification performance has been greatly improved.Deep learning-based classification networks are generally established through large neural networks and require a large number of training samples to optimize their weights.However,it is difficult and expensive to collect vast amounts of heart sound data in reality,which leads to the difficulty in obtaining sufficient heart sound training samples.Therefore,the heart sound classification faces the challenge of small samples,and there is no effective solution at present.For the heart sound classification in the case of small samples,this paper mainly carries out the following works.(1)A heart sound classification method based on supervised threshold is proposed to solve the issue that estimating the label of original heart sound sample caused by the current training data augmentation method through the segmentation of original heart sound sample.In the current mainstream heart sound classification methods based on deep learning networks,the original heart sound training samples are first segmentded into a few of fragments,and then each fragment is used to train classification network,so as to realize training data augmentation,which is the main method to solve the shortage of heart sound training samples.However,this leads to the issue that how to estimate the label of original sample depend on the prediction results of the fragments in the prediction process.The traditional average method of fragments'classification probabilities and the voting method of fragments' prediction labels are not very effective in solving this issue.Therefore,we propose a heart sound classification method based on supervised threshold.Its key innovation is to design a rule for estimating the label of original heart sound sample based on the supervised threshold.The rule is used to estimate the label of original heart sound sample depend on the fragments' prediction labels,and its design idea can be extended to solve this kind of problem,that is,how to estimate the sample's category according to the local category of the sample.Compared with baseline,the proposed heart sound classification method derived from the combination of baseline and this rule significantly improved the classification effect.(2)Establishes a classification network that directly links the data dimensionality reduction process and the process of classifying using low-dimensional representations,which can realize the joint optimization of these two processes.And further,a novel heart sound classification method is proposed based on this network.Heart sound samples are not only difficult to obtain,but also have high data dimension.Therefore,heart sound classification is faced with the problem of high-dimensional data classification in the case of small samples.Traditionally,the main method to solve this problem is to first reduce the data dimension and then use the low-dimensional representations to classify.In this type of methods,the process of data dimension reduction and the process of performing classification tasks using low-dimensional representations are generally independently,namely the dimension reduction is not constrained by classifying.This can easily lead to data information with category attributes being lost too much in the dimensionality reduction process,resulting in the discrimination of low-dimensional representations is insufficient,which brings the adverse effect for the classification task.To this end,this paper selects semi-nonnegative matrix factorization(Semi-NMF)to perform data dimensionality reduction,and establishes a classification network which can jointly optimize the process of data dimensionality reduction and the process of classifying using low-dimensional representations.This can promote dimensionality reduction to be performed in a direction that is beneficial for classification task,so that the data information containing category attributes can be well retained in the low-dimensional representations in the process of dimensionality reduction,so as to make the low-dimensional representations more differentiated.A novel heart sound classification method is proposed based on this network.Compared with the heart sound classification method,whose dimensionality reduction process and classification process are carried out independently,the proposed method achieves higher classification accuracy.In addition,compared with the deep learning-based heart sound classification methods,the novel method also obtains significantly better classification performance in the case of small samples.(3)A mathematical model of sparse Semi-NMF is proposed,which is used to improve the classification network recommended in(2),and thus to derive a new classification network that directly connecting the process of sparse dimensionality reduction and the process of classifying using sparse low-dimensional representations.And further,a novel heart sound classification method is proposed based on this network.Combining sparse constraints with data dimensionality reduction to obtain sparse low-dimensional representations has been proved helpful in making low-dimensional representations more differentiated,and thus achieving better classification effect.In this type of methods,the parameter that controls the sparseness of low-dimensional representations should be specified in advance,and the enumeration method is usually used to search the appropriate sparse parameter.However,enumeration method is very time-consuming,and the searched sparse parameter is not necessarily optimal.We propose the sparse Semi-NMF model,which is used for sparse dimension reduction,and establishe a new classification network that directly links the sparse dimensional-reduction process and the process of classifying using sparse low-dimensional representations.In the novel network,the sparse parameter of sparse Semi-NMF is weight,which can be optimized by learning.Therefore,it realizes the adaptive sparsity of factors.This network is applied to the heart sound classification,which further improves the classification effect and better solves the problem of heart sound classification in the case of small samples.To sum up,this paper focuses the problem of heart sound classification in the case of small samples.First of all,we studied the lack of the means used by mainstream deep learning methods to solve this problem,and proposed a heart sound classification method based on the supervised threshold,which significantly improved the classification effect of deep learning method.Then,two kinds of classification networks are designed to solve the problem.And two novel heart sound classification methods are proposed based on these two networks.In the case of small samples,the novel methods have achieved better classification effect compared with the excellent methods in recent years.Finally,this paper summarizes the contents and innovations,and points out some problems or directions for further research.