Research On Heart Sound Classification Algorithm Based On Supervised Machine Learning

The heart sound signal can detect the working condition of the heart and arterial vessels,and the analysis results can be used as an auxiliary diagnostic tool for predicting cardiovascular diseases.Currently,the main means of performing heart sound detection is heart sound auscultation based entirely on physician experience;therefore,in recent years,computer-aided detection techniques is heavily applied to quantitative acquisition and analysis of heart sound signals,which are important for noninvasive adjunctive diagnostic studies and early screening of cardiovascular diseases.In this paper,we conduct research based on a supervised machine learning heart sound classification algorithm,and the related work is summarized as follows.(1)According to the characteristics of heart sound signals,a comprehensive noise reduction method with Butterworth bandpass filter and wavelet noise reduction is proposed.Firstly,a sixthorder bandpass filter with a passband frequency of 25Hz-400 Hz is set to remove low-frequency artifacts,baseline drift,and high-frequency interference from the original heart sound signal.An adaptive threshold heart sound noise reduction method is then proposed,according to the fact that the duration of the first heart sound and the second heart sound in one heartbeat cycle is generally less than 25% of the total time,the value of 75% of the wavelet coefficient sorting on each scale is set as the filtering parameter,and the nonlinear threshold function is determined by combining parameters such as mean and variance to obtain better coefficients to process the wavelets and achieve better noise reduction performance.Subsequently,pre-processing operations such as downsampling,normalization,and data interception are performed to reduce the complexity of feature calculation while the number of training heart sound samples can be increased.(2)This paper extracts heart sound kinetic features based on no segmentation,proposes an improved MFCC feature for feature extraction of heart sound signal classification,and proposes a support vector machine-based heart sound classification algorithm.The difference coefficients of MFCC features are extracted,the optimal dimensionality of MFCC features and the impact of introducing difference coefficients on the performance of the classification model are experimentally explored,and the MFCC features of the SVM-based heart sound classification model used in this paper are finally determined: 14-dimensional MFCC features + 14-dimensional MFCC third-order difference,which achieves a classification accuracy of 99.04%,an accuracy of99.03%,sensitivity and specificity of 98.81% and 99.26%,respectively,and classification performance of 99.03% for F1 mean on the Yaseen2018 dataset.(3)A recurrent neural network-based heart sound classification algorithm is proposed.Four recurrent neural networks with different network structures are proposed to explore the classification performance of the improved MFCC features on these four recurrent neural networks,and it is concluded through experiments that the best classification recognition effect is achieved with the structure of the recurrent neural network set as 2-layer GRU plus 1-layer LSTM,with an accuracy of 95.77%,sensitivity and specificity of 97.02% and 94.51%,the accuracy of 96.95%,and F1 mean of 95.82% on the Challenge 2016 dataset.Two neural network models,CNN and CNN-LSTM,were also constructed for heart sound classification comparison experiments,and the experimental results show that the RNN-d network model constructed in this paper has the best classification performance.To further illustrate the performance of the model proposed in this paper,the heart sound classification network model proposed in this paper is compared with other classification algorithms proposed in the literature and typical machine learning classification algorithms to further demonstrate the superiority of the classification algorithm in this paper.