Research On Heart Sound Classification Based On Machine Learning

Cardiovascular and cerebrovascular diseases have been the biggest killers of mankind in the past ten years.In 2018,they caused more than 17 million deaths worldwide.The number of patients in China has reached more than 200 million.Heart sounds are important physiological signals of heart health.Therefore,it is of great significance to study the automatic recognition algorithm of heart sound signals.Machine learning has performed well in classification problems in recent years.This paper combines some methods of signal processing and machine learning to analyze the problem of distinguishing normal and abnormal heart sound signals.Comparing different machine learning algorithm models,a feasible model is proposedThe main work and results of this article are as follows:(1)Pre-process the heart sound data.Wavelet threshold denoising is used to filter out the noise in the signal,and a hidden semi-Markov model is used to find the position of S1,systole,S2,and diastole in cardiac cycle,and then extracting 96 features,including features in time-frequency domain and signal envelope on different frequency bands.T-test was performed on the selected 96 features,and 77 features were selected to prepare for the next step of training traditional machine learning models.(2)Three traditional machine learning algorithms(support vector machine,random forest and K-nearest neighbor)are used to model on the dataset.Three parameter selection experiments are designed to select the optimal parameter combination.This article conducted a 10-fold cross-validation experiment on the dataset,and compared the advantages and disadvantages of three models.Then SMOTE algorithm is studied to solve the problem of imbalance of the feature matrix.(3)Two types of CNN models were constructed—single CNN model and dual CNN model.First,the modified frequency slice wavelet transform is used to convert the one-dimensional signal into two-dimensional time-frequency image.The optimal duration parameters of the cardiac cycle window are studied.The single CNN model and the dual CNN model is designed to classify heart sounds.Similarities and differences between models are also studied.The sensitivity,specificity,accuracy and measure of accuracy of the dual CNN model are 0.95,0.93,0.93 and 0.94,respectively.It proves that the model can classify heart sounds efficiently.