Study Of Heart Failure Staging Using Heart Sound Entropy

Recently,cardiovascular disease had become the biggest killer affecting the people's health,and heart failure(HF)is a complicated clinical syndrome manifested in the terminal stage of various types of cardiovascular diseases with a high mortality and morbidity rate.The American College of Cardiology and American Heart Association(ACC/AHA)divide heart failure into four stages(namely stage A,stage B,stage C and stage D)based on the cardiac structural and functional changes and the HF symptoms.Diagnosis,treatment and prevention of heart failure patients with different stages can effectively improve the prognosis of patients and delay the deterioration of heart failure.However,the current clinical detection of heart failure still has limitations with lagging diagnosis and high-cost,and there is no objective and accurate means of heart failure staging.As one of the valuable physiological signal,heart sound can reflect the mechanical movement of the large vein and cardiac system in time,and there are some connection between the heart sound and heart failure staging for that heart sound would be distorted before the occurrence of HF symptoms and electro cardiac abnormality in patients.Therefore,the analysis and detection of heart sound can provide an effective auxiliary detection method for the heart failure staging in clinical.On the in-depth study of the connection between heart sound and heart failure,this paper extracted the entropy features based on the time-frequency characteristics and nonlinear characteristics of heart sounds,and selected feature vectors by statistical analysis to realize the recognition of heart failure staging.First,this paper introduces the basic knowledge of heart sound,and explains the relationship between heart sound and heart failure staging in detail.Since the heart sound is an unstable time-frequency signal that is susceptible to interference,the signal pre-processing was implemented by down-sampling,normalization and wavelet denoising.Four sets of experiment were used to obtain the best wavelet parameters by using the single variable method,including decomposition layer,threshold function,wavelet basis and threshold.After the preprocessing,in order to reflect the complexity distribution of heart sound in the frequency domain,this paper combined wavelet packet decomposition with Shannon information entropy to extract the entropy feature of frequency domain from three aspects of energy,singular value and power spectrum,including wavelet packet energy entropy(WPEE),wavelet packet singular entropy(WPSE)and sub-band power spectrum entropy(SPSE1~SPSE8).The results show that the WPEE,WPSE and SPSE1 gradually increases during stage A to stage D,indicating that the randomness of heart sound distribution in the frequency domain increases with the development of heart failure,and the information of frequency domain was more chaotic.Moreover,the sample entropy(SampEn)was extracted to reflect the complexity of heart sound time series.Result shows it is opposite to the trend of WPEE,indicating that the degree of disorder in time domain is gradually decreasing in the development of heart failure.Through statistical analysis,the WPEE,WPSE and SPSE1 have significant differences among normal and different heart failure stages,indicating that these features can effectively embody the samples' difference of patients with different stages,but not the rest.In order to realize the recognition of heart failure staging,this paper selected WPEE,WPSE and SPSE1,which has differences in each group,as feature vector to input into artificial neural network,K-nearest neighbor,decision tree and support vector machine(SVM),and adjust the classifier parameters to get the best classification effect.Then,we validated our method with a dataset of 280 recorded HS signals consisting of 80 normal signals,48 of stage A,48 of stage B,56 of stage C and 48 of stage D,respectively.Results show that the staging accuracy of SVM employing radial basis kernel function is the best of 89.29%.Since the penalty factor and kernel function parameters are the key factors affecting the SVM classification performance,this paper employed genetic algorithm(GA)and particle swarm optimization(PSO)for parameter optimization instead of the exhaustive method with time-consuming and low-accuracy.The result shows that the SVM parameter optimization based on GA obtained the highest accuracy of 91.43%,indicating that it is better than PSO-SVM.