Cardiorespiratory Sound Separation Based On Regression Transfer NMF And Densely Connected Networks

Cardiorespiratory sound signals are rich in physiological and pathological information that characterizes the health of the cardiopulmonary system.Clinically,doctors diagnose the patient's cardiopulmonary system by listening to and analyzing the heart sounds and lung sound signals collected by the stethoscope.However,the mutual aliasing of heart sounds and lung sounds interferes with the doctor's auscultation and reduces the effect of auscultation.In order to solve this problem,in summarizing the basis of previous research results,this paper starts with non-negative matrix factorization and neural network technology,proposes two new methods of cardiorespiratory sound separation and validates them through simulation experiments.1)There are individual differences in cardiorespiratory sound signals collected from different subjects.If the cardiorespiratory sound separation of target subject is obtained by using the supervision information provided by the heart sound and lung sound signals of other subjects,the separation effect is reduced by the influence of individual differences.In order to solve this problem,this paper draws on the idea of transfer learning,and proposes a method of cardiorespiratory sound separation based on regression transfer non-negative matrix factorization.We firstly mined the frequency domain information and time domain recursive information common to the cardiorespiratory sound signals of different subjects.And then,the common frequency domain information and the time domain recursive information of the cross-subjects were learned from the cardiorespiratory sound signals of other subjects and were incorporated into the non-negative matrix factorization of the spectrum of the target cardiorespiratory sound signal,which helps the learning of the spectral components of the target heart sound and lung sound.After that,we used time domain k-means clustering to determine the time-spectrum components of the heart sounds and lung sounds.Finally the time-domain signals of the heart sounds and lung sounds are reconstructed by the time-frequency mask and the inverse short-time Fourier transform.We evaluated the cardiorespiratory sound separation performance of the proposed method on the self-constructed dataset and clinical dataset.The experimental results show that the cardiorespiratory sound separation method based on regression transfer non-negative matrix factorization has better cardiorespiratory sound separation performance than the existing non-negative matrix factorization based cardiorespiratory sound separation method,and can be effectively applied to the clinical cardiorespiratory sound signals.2)The cardiorespiratory sound separation methods based on non-negative matrix factorization assume that the cardiorespiratory sound is linearly aliased in the time-frequency domain.However,the internal structure of the thoracic cavity is complex,so the heart sound and lung sound components collected by the stethoscope on the body surface may have nonlinear aliasing in the time-frequency domain.To this end,we designed a densely connected LSTM network to achieve cardiorespiratory separation,using the LSTM network to handle nonlinear aliasing of cardiorespiratory sound components,and to enhance the separation performance by capturing the temporal correlation of cardiorespiratory sound components.In order to optimize the network,the information flow and the convergence speed of the network are improved by adopting a densely connected network structure.Affected by individual differences in cardiorespiratory sounds,there are mean and variance shifts in the distribution of hidden layer features on the training set and test set.In order to solve this problem,we use batch normalization to zero-mean and unit variance normalization of network hidden layer features to overcome the mean and variance shifts,thus improving the generalization ability of the network.In summary,this paper proposes a cardiorespiratory sound separation method based on batch normalization and densely connected LSTM networks.The simulation results on the self-constructed dataset show that the proposed method is feasible in the cardiorespiratory sound separation task,which has obvious advantages compared with the existing linear separation method.