Research On Heart Sound Feature Extraction And Recognition For Prior Conditions Of Different Cardiac Cycle Structures

The heart sound signal is a physiological signal generated by the heart movement,and it contains information about the cardiovascular health status of the human body that is irreplaceable by the electrocardiogram.Through the analysis and learning of heart sound signals,we can provide the decision-making basis for the diagnosis and monitoring of the heart diseases which has important social and economic benefits.Specifically,the heart sound signal is a quasi-periodic signal composed of a series of cardiac cycles,in which the heart murmurs reflecting different types of pathological information usually appear at different positions of the cardiac cycles.Therefore,under the condition of the known cardiac cycle structures,it is often possible to extract effective features to perform heart sound classification according to the timing structures of the signal.However,the accurate segmentation of the cardiac cycle structures within the heart sound signal is difficult and costly.Sometimes,the heart sound feature extraction and recognition research can only be carried out under the unknown condition of the cardiac cycle structures.Meanwhile,in many application scenarios,we only pay attention to whether the heart sound signal is abnormal and do not distinguish the specific abnormal types.Consequently,such recognition task can also be completed without the segmentation information of the cardiac cycle structures.Therefore,it is also necessary to study the abnormal heart sound recognition problem under the unknown condition of the cardiac cycle structures.In recent years,a large number of scholars have carried out related research works on the methods of extracting and identifying heart sound features under the two different conditions of known and unknown cardiac cycle structures.Although some progress has been made in these research works,there are still many problems to be solved due to the complexity of the heart sound signals.In this paper,we conduct the studies against the problems in the current research,including lacking the distinguishing and structural heart sound features for time aligned signals under the known condition of the cardiac cycle structures,and that the extracted features are difficult to match for signals susceptible to the time shifting under the unknown condition of the cardiac cycle structures.The main research contents and innovations are as follows:(1)Under the known condition of the cardiac cycle structures,a discriminative feature extraction method for effectively utilizing the sample category information is proposed for the problem of poor discriminative heart sound features extracted by the existing methods.More separable features are obtained by using the partial least squares method to find the projection directions that maximize the correlation between the cardiac cycles and category labels while maximizing the variance within the cardiac cycles.Further considering the case that parts of the obtained features are linearly inseparable,the original features are mapped into a high-dimensional linearly separable kernel space.In the kernel space,the kernel partial least squares method is used to increase the distinguishability of the extracted features.The experimental results show that the proposed partial least squares feature extraction method has better classification performance,and the classification performance can be further improved by the kernel method.(2)Under the known condition of the cardiac cycle structures,aiming at the problem that the traditional heart sound features have poor ability to express the time-frequency structures of the heart sound components and their relative position information whitin the cardiac cycle,a feature extraction method for maintaining the complete time-frequency structure of the cardiac cycle is proposed.By using the tensor decomposition method,the features of the spectrogram in the cardiac cycle structure are simultaneously dimensionreduced in the time domain and the frequency domain directions,thereby obtaining the features with better representation ability to maintain their time-frequency structure.Further,the heart sound category information is introduced by the Fisher discriminant criterion and the partial least squares criterion in the tensor decomposition process to increase the distinguishability of the extracted features.The experimental results show that the proposed tensor decomposition feature extraction method has better classification performance,and the Fisher discriminant criterion or partial least squares criterion can further improve the classification performance.(3)Under the unknown condition of the cardiac cycle structures,aiming at the problem that the existing time-shift invariant heart sound feature extraction methods cannot reflect the characteristics of each subband of the heart sound signals and their autocorrelation in each order finely due to the lack of the in-depth correlation analysis between the signal autocorrelation information,a feature extraction method for abnormal heart sound recognition based on the time delay domain dependencies of the subband autocorrelation features is proposed.By using the average amplitude difference function for each subband of the heart sound signal,the subband autocorrelation features having certain long and short-term dependency relation in the time delay domain are obtained.The long and short-term memory network is further used to model this dependency relation,and the subband autocorrelation modeling features that reflect the characteristics of the heart sound signal finely are extracted.The experimental results show that the proposed modeling features have better abnormal heart sound recognition performance.(4)Under the unknown condition of the cardiac cycle structures,aiming at the problem that the previous heart sound feature extraction methods cannot take into account both the local important information and the time-shift invariant ability,a time-shift invariant local time-frequency structure feature extraction method for abnormal heart sound recognition is proposed.By using the convolutional neural network,the local features of the heart sound spectrogram are extracted layer by layer.The temporal max pooling method is used for the final convolutional layer to obtain the feature representation,thereby eliminating the influence of the time shift while preserving the most important local information.The experimental results show that the proposed local time-frequency structure feature extraction method has better abnormal heart sound recognition performance.