Research Of Heart Failure Typing Based On Recurrent Neural Networks

Heart failure(HF)is increasingly becoming one of the major threats to human health duo to the high morbidity and mortality,especially with the aging of population and the improvement of living standard.According to the left ventricular ejection fraction,HF can be divided into two categories — HF with preserved ejection fraction(HFp EF)and HF with reduced ejection fraction(HFr EF).Some studies have shown that different categories of HF differ greatly in the pathogenesis and treatment plans.Therefore,the earlier the accurate diagnosis and treatment the patients get,the more efficiently delay of the HF progress could achieve;however,there is a serious lag in the clinical diagnosis of HF.Heart sound(HS)can reflect the mechanical dysfunctions of myocardial activity directly before the occurrence of the cardiac organic lesion,which is a non-stationary physiological signal of cardiovascular system produced by the beat of muscles.HFp EF and HFr EF have differences in the mechanism of ventricular remodeling,which also reflects on HS.Therefore,HS can be used for early diagnosis of cardiovascular diseases such as HF.In view of the intrinsic connection between HS and HF,as well as the characteristics of deep learning,this paper proposed a new method of HF typing based on recurrent neural networks,which includes the following aspects:(1)Get the HS data set needed for deep learning,which includes the HS of normal,HFp EF and HFr EF.The normal HS was obtained from the public database,and the HS of HF patients came from the independent collection.To meet the requirements of the deep learning models,the original HS needs to be preprocessed,including de-noising,segmentation,resample and normalization.An improved threshold function and estimation method was used in wavelet de-noising.The strategy of fixed starting point and length was adopted to obtain HS frames in this paper: first,logistic regression-based hidden semi-Markov model is selected to detect the onset of S1,and then based on this point,an appropriate fixed length was selected to segment HS.(2)The automatic HF typing methods based on improved recurrent neural networks,including long short-term memory(LSTM)and gated recurrent unit(GRU),were proposed.Taking HS frames as the input of networks directly without the step of hand-crafted feature extraction and selection,as the relevant recurrent neural networks are good at mining the deep features from time series signals.Then fed the deep features into the softmax classifier to realize the classification of normal,HFp EF and HFr EF HS signals.The results of ten-fold cross validation show that the performance of GRU model is competitive with the LSTM model,with the average accuracy of 96.60%.In addition,to compare with the performance of deep learning and traditional machine learning,we extracted multiple types of HS entropy as the inputs of support vector machine(SVM).The results show that the deep learning models outperform the SVM in HF classification.As a representative of traditional knowledge-driven methods,the unsatisfactory results of SVM may be related to the selection of features.Moreover,the impact of the length of HS frames on HF classification results based on GRU model was studied.The results show that the frame length was related to the amount of information contained,and when the frame length set to 1.6 s,the GRU model achieves the best performance.(3)A time-space advanced feature fusion model combining the GRU and fully convolutional networks(FCN)was proposed.The FCN is used for space feature extraction from time series,while the GRU is able to learn the temporal dependencies within the time series.The GRU-FCN can reach the average accuracy to 97.60%,which is 0.7% higher than that of GRU model,indicating that the fusion model can better distinguish HS of normal,HFp EF and HFr EF,which could provide a strong reference for the HF early diagnosis.