Research On Heart Sound Classification Based On Automatic Analysis Technology

Heart sound signal analysis plays an important role in the clinical detection of cardiovascular diseases and physical examination.At present,the most widely used heart sound signal analysis method is automatic analysis technology,it mainly includes core steps of denoising,segmentation,feature extraction and classification.In this paper,the automatic analysis technology of heart sound is analyzed,and the existing difficulties are studied.The main work can be summarized as follows:(1)Among the denoising algorithms,this paper studies the wavelet-based denoising method,and proposes an adaptive wavelet denoising method based on coif-5.However,the noise has nonstationary,chaos and other natures,the theory that the wavelet coefficient of heart sound is greater than the noise is not completely established.And the wavelet method cannot track the change of the noise,and the threshold parameters extracted in the long-term collection of heart sounds are prone to errors.Therefore,a new method of heart sound denoising is proposed,which combining the improved minimum control recursive average and the optimal modified logarithmic spectral amplitude estimation.It uses a short-time window for smooth dynamic tracking and estimation of the minimum noise value,and minimize noise by minimizing the difference between clean heart sounds and estimated clean heart sounds.Experimental results verify the effectiveness of the proposed algorithm.(2)In this work,we study the segmentation methods based on threshold and hidden Markov model.The feature of threshold segmentation method is low complexity,and it is more suitable for high real-time processing system.In this study,we propose a new envelope extraction method based on the principle of non-stationary system identification.It is based on the feature that the short-term stationary of the heart sound is stronger than the short-term stationary of the noise,and sequentially recognizes the non-stationary correlation of the heart sound frame,thereby obtaining a more effective feature envelope.However,the segmentation accuracy of threshold segmentation method decreases with the decrease of signal-to-noise ratio.So that we studied the personalized Gaussian mixture model and duration-dependent hidden Markov model,and proposed to constrain the hidden Markov model by Gaussian mixture model.In order to avoid the error of segmentation components caused by fuzzy time-domain features,convolution neural network is used to classify the optimized Mel frequency cepstral coefficient features of fundamental heart sounds,so that the time periods of S1 and S2 are distinguished first,and then the rest components of heart cycle are defined.Finally,the proposed algorithm achieves 92.93% segmentation accuracy under the test data set of challenge 2016,and it’s results are better than other algorithms at this stage.(3)Heart sound classification is the ultimate goal of the analysis system,and we mainly studied the classification of normal and abnormal heart sounds.In this paper,a support vector machine heart sound classification method based on fused multi-features is proposed,which uses nonlinear radial basis functions as kernel functions.In this study,we used the Challenge 2016 data set to construct the training set and test set,and used the uploaded data from the Whale community to construct the verification set,and completed the heart sound classification through a support vector machine based on non-linear radial basis function.In this process,the 420-dimensional features such as time domain duration,amplitude,energy,spectrum,kurtosis of each heart sound record are extracted,and 271-dimensional features are obtained by principal component analysis.The accuracy of the test set is 82.39%,the accuracy of the verification set is 80.44%,and the overall score of sensitivity and specificity of the test set is 0.8629.