Research On Heart Sound Classification Methods Based On Hybrid Neural Networks

Cardiovascular disease is one of the most important health problems worldwide.Early diagnosis of cardiovascular disease is essential for the treatment and prevention of this disease.Heart sound signals reflect the health status of the heart.With the rapid development of machine learning,heart sound automatic analysis technology has gradually become a research hotspot.Automatic classification and recognition of heart sounds can provide a more accurate and fast inspection method.Therefore,the study of heart sound classification is of great significance for the diagnosis of early cardiovascular diseases.This study uses neural networks to classify normal and abnormal heart sound signals.The main work is as follows :(1)Considering the time domain and frequency domain information contained in the heart sound signal,the frequency domain features are extracted for the preprocessed heart sound signal,and a classification method based on CNN and LSTM hybrid model CRNN(Convolutional Recurrent Neural Network)is proposed.Firstly,the large receptive field space feature extraction module is connected,and the convolution operation adopts the hollow convolution method to increase the receptive field of the convolution layer.At the same time,the residual connection is added to avoid the problem of gradient disappearance.Then input the time information extraction module,which contains the long-term and short-term memory layer,which can extract its time domain information,and achieve better results than the model using a single convolutional neural network or a long-term and short-term memory network.The comparative experiment proves that the residual and dilated convolution can improve the classification effect of the model.(2)Aiming at the problem that the current heart sound classification algorithm uses a single model,a heart sound classification method based on time-frequency component attention hybrid model ARNN(Attention Recurrent Neural Network)is proposed.The model uses the time-frequency component attention layer instead of the convolution layer to extract features by calculating the attention of the sub-vector features in the time domain and the sub-vector features in the frequency domain.At the same time,the long-term and short-term memory layers are used to extract features in parallel,and the extracted features are fused for classification.Using multiple data sets to fuse,while balancing the proportion of positive and negative samples,the classification effect is further improved.Using multi-source input,the heart sound sequence is used as input one,the long-term and short-term memory layer is used to extract the time series features,and the spectrogram is used as input two.The frequency domain features extracted by attention calculation are fused,and good classification results are obtained.Experiments using a variety of audio features have achieved good classification results.The hybrid model proposed in this study achieved an accuracy of 90.57 % on public datasets.After using the self-attention layer instead of the convolution layer to improve the model,the average accuracy rate on the data set was more than 91 %,and the effect was significantly improved.After expanding the data and balancing the positive and negative samples,a variety of features were used to test the effectiveness of the model.The highest accuracy,sensitivity and specificity were 95.5 %,95.7 % and 95.3 %,respectively,which were better than the best method of comparison 0.7 %,3 % and 0.4 %,and the indicators were very balanced.The results show that the heart sound classification model proposed in this study can achieve better results,which proves the feasibility of the attention mechanism in the heart sound classification task.