Research And Application Of Convolutional Neural Network In Classification In Heart Sounds

In recent years,the living standards of the Chinese people have gradually improved,and many people's living habits have begun to develop in an unhealthy direction,which has led to a continuous increase in the incidence of cardiovascular disease in my country.Through the auscultation of heart sounds in the primary health care examination in my country,if aortic stenosis can be detected,the patient can be classified as a suspected case of cardiovascular disease at the initial stage of onset.On this basis,the preventive treatment of cardiovascular disease can be carried out early.It can improve the patient's informed rate and cure rate.my country has a vast territory and relatively short medical resources.The use of a doctor-patient one-to-one model for heart sound auscultation consumes medical resources.In addition,the results of auscultation will also be affected by the doctor's experience and subjective judgment.Therefore,the use of computers to achieve assisted heart sound auscultation for heart early screening for vascular diseases is very necessary.Traditional heart sound classification methods are highly dependent on the precise segmentation of heart sound signals and the effectiveness of the extracted features.In addition,previous studies often only use heart sound data sets that are small,carefully selected,and noisy.Even if these algorithms can achieve a high accuracy,there are still problems that are not conducive to popularization and real-time decision-making.Although convolutional neural networks have made their debut in the field of heart sound classification,most of the input data of heart sound classification models rely on accurate heart sound segmentation operations.The training model is based on a small number of samples.The model itself has a single structure and cannot fully mine and extract heart sound signals features,which resulting in the classification performance still needs to be improved.Based on the above problems,we intend to use a large amount of heart sound data to train a heart sound classification model based on deep learning,and explore a new and practical method for heart sound classification that does not require precise segmentation to improve the performance of heart sound classification.On the other hand,the spatial structure of the deep network model should ensure the excellent classification effect,and the model parameters should be as few as possible,so as to provide some new ideas of the heart sound classification algorithm for the application of the heart sound classification algorithm in the portable auscultation equipment or the server of the telemedicine center.Therefore,the topic of our article has launched a research on the application of convolutional neural networks in the classification of heart sound signals.The main contributions and innovations of our article include:(1)A heart sound data preprocessing model combining sliding window and Mel frequency coefficient is proposed.The model not only does not require precise segmentation of the original heart sounds,but also alleviates the problem of small training costs to a certain extent.The model first uses a sliding window to slice the original heart sound data to obtain a large number of heart sound subsequences.On this basis,the one-dimensional heart sound sub-sequence is converted into a two-dimensional time-frequency matrix using Mel frequency coefficients,and a large number of feature matrices that can be used to train deep convolutional neural networks are obtained.The preprocessing model is used in the experiment to expand the Challenge 2016 data set containing 3240 heart sound recordings to14,692 positive heart sound subsequences and 14692 negative heart sound subsequences,and also expand the CHSC data set containing 585 heart sound recordings to 3314 positive heart sound subsequences and 3312 negative heart sound subsequences.It provides a more comprehensive and reliable data foundation for the subsequent heart sound classification model based on deep learning.(2)A convolutional neural network architecture based on fusion of global information is proposed for heart sound classification.The network architecture is dedicated to improving the performance of heart sound classification.We design a heart sound classification model from the perspective of a deep convolutional neural network.At the same time,multiple global pooling layers are used to extract the characteristic information of the neural network from low-level to high-level,and integrate it into the final decision-making layer of the network.We compare various high-level and low-level feature information fusion strategies through a large number of experiments,and show that the heart sound classification model proposed by us can improve the accuracy of 0.08%?2.57%,the sensitivity of 0.02%?1.85%,and the specificity of 0.09%? 1.65%.It is verified that the architecture proposed by us can effectively improve the effect of heart sound classification.(3)Three types of heart sound classification models based on deep convolutional neural networks are constructed.In order to find a more suitable module structure for heart sound classification,we summarize three convolutional neural network module structures based on the feature information transfer method between convolutional layers: a convolution module based on simple connections,a convolution module based on jump connections,and a convolution module based on dense connections.These convolution modules and the previous network architecture respectively construct three types of heart sound classification models based on deep convolutional neural networks.Experiments show that the convolutional neural network model based on dense connections is more suitable for heart sound classification,and improve the accuracy of 2.49%?3.70%,the sensitivity of 0.92%?2.47%,and the specificity of1.73%?3.84%.(4)Three lightweight modules for channel enhancement based on squeeze-stimulus mechanism and depth separable convolution are constructed.In view of the large number of parameters of the convolutional neural network model based on dense connections,this paper respectively uses squeeze excitation-deep convolution-point convolution module,deep convolution-squeeze excitation-point convolution module,deep convolution-point convolution-squeeze excitation module to improve the convolutional neural network model based on dense connections.Compared with the heart sound classification model based on dense connections before the improvement,under the premise of ensuring the classification effect,the parameter amount of the improved model is reduced by 44.11% and achieve the purpose of lightweight model.In order to verify the effectiveness of the heart sound classification model based on the deep convolutional network model constructed in this article,the Challenge 2016 data set and the CHSC data set are used to train and test all the heart sound classification models in this article.The experimental results show that the convolutional neural network architecture based on global information fusion proposed in our paper combined with the densely connected convolution module to build a densely connected heart sound classification model can achieve better classification results;using deep convolution-squeeze The compression excitation-point convolution module improves the model,and can obtain a lightweight heart sound classification model with better classification effect,which provided a new idea for the ultimate realization of computer-assisted auscultation and cardiovascular disease screening.