A Study On The Identification Of Heart Failure With Preserved Ejection Fraction Based On Heart Sound And Its Multifractal Characteristics Analysis

Heart failure with preserved ejection fraction(HFpEF)is a complex and heterogeneous clinical syndrome,and accounts for approximately 50% of all heart failure patients.The pathophysiological mechanism of HFpEF is not completely clear and its clinical diagnosis is difficult.Heart sound can provide effective information for early recognition of cardiac abnormalities due to its direct reflection of the mechanical properties of cardiac activity.In this paper,two automatic recognition methods of HFpEF were proposed based on heart sound signal from the traditional features extraction based method and deep learning respectively,aiming to explore a non-invasive and efficient method as well as a new direction for HFpEF diagnosis.The main contents are as follows:Firstly,a new threshold function was proposed to overcome the shortcomings of traditional soft and hard threshold functions,and a new wavelet denoising scheme without setting any hyper-parameters was established.In this method,different mother wavelets as well as decomposition layers for different signals can be selected adaptively,and by comparing the signal to noise ratio,mean square error and correlation coefficient of different thresholds and threshold functions at different noise intensity,we found that the unified threshold and the proposed new threshold function can achieve better denoising effect,which can not only remove the redundant noises but also retain the pathological heart murmurs well.Secondly,the nine multifractal features consisting of generalized Hurst exponent,Renyi index and multifractal spectrum related parameters of heart sound signal were extracted and analyzed comprehensively based on multifractal detrended fluctuation analysis.The results show that the heart sound signals are long-range anti-correlated,and heart sounds of HFpEF patients have a smaller anti-correlation than that of healthy people;Murmurs may generated in heart sound signals of HFpEF patients,and the cardiac contractility of HFpEF patients may be impaired,but there is no significant differences between HFpEF patients and healthy people;In addition,the multifractal intensity,complexity and volatility of HFpEF are all drecreased.Thirdly,the research on HFpEF recognition based on traditional machine learning was carried out.Principal component analysis was used to reduce the dimension of the multifractal features with significant differences between heart sound of healthy people and HFpEF patients.The accuracy,sensitivity and specificity were combined as the evaluation criterions in order to choose the optimal input and activation function.Then,the performance of optimally pruned eatremely learning machine(OP-ELM),eatremely learning machine(ELM)with its number of hidden neurons selected by grid search algorithm,as well as support vector machine(SVM)whose hyper-parameter values were selected by particle swarm optimization algorithm,was compared and analyzed.The results show that the performance of both OP-ELM and ELM is significantly improved compared with that of SVM.Due to the influence of the optimal pruning process,the 10-fold cross-verification time of OP-ELM is longer than that of ELM,but its average accuracy,sensitivity and specificity are improved by 2.7%,5% and 0.42%,respectively,and the recognition rate of HFpEF heart sounds by OP-ELM reaches92.27%.In addition,although the speed of OP-ELM is reduced,it does not have any impact on the practical application.Fourthly,a novel approach for HFpEF identification was proposed based on deep learning,that is,the deep features of heart sounds were automatically extracted based on Le Net5,which is a classicifical model of convolutional neural network,and the classification was achieved by OP-ELM classifier.The results show that the combination of OP-ELM classifier and deep learning model can improve the model’s ability to recognize HFpEF heart sounds,and its performance is better than that of the method combined with SVM.However,the performance of OP-ELM combined with Le Net5 is worse than that of OP-ELM combined with multi-fractal features due to the simplicity of Le Net5 network and its insufficient ability to extract deep features of heart sounds.