Research On Classification Algorithm Of Heart Sound Without Segmentation

Heart sound was an important physiological signal in human body.In most cases,heart sound signal needed to be segmented to achieve feature extraction.However,pathological heart sound was more complex than normal heart sound,and it was difficult to achieve segmentation for complex pathological heart sound.To solve this problem,two kinds of heart sound feature extraction method without segmentation were proposed.(1)Heart sound was a quasi-periodic and chaotic signal,and recurrence plot was an important method to analyze periodicity and chaos of time series.Firstly,equal-part and equal-length processing was used on heart sound signal,which not only extracted the frequency band of heart sound,but also reduced the length of data,as well as avoided the problem of too short data length reduction.Secondly,an adaptive threshold acquisition method was proposed,and considered the complexity of heart sound signal,single threshold was difficult to ensure that heart sound feature was effectively presented in recurrence plot.Therefore,multi-threshold fusion heart sound recurrence plot was constructed,which made time-domain features of heart sound to be highlighted and enlarged.Finally two-dimensional features were extracted from it.For this kind of feature extraction method,heart sound database of research group and an public heart sound database were used for analysis,and when support vector machine was used for classification,the recognition rate could reach to 90%.However,this feature extraction algorithm had two disadvantages:First,the time to find adaptive threshold was long,and the efficiency was too low when analyzing large amounts of data.Second,for complex pathological heart sound analysis,the result was not good,the applicability was not wide enough.So the second kind of feature extraction method was proposed.(2)Heart sound was a low-frequency and narrow-band signal,and noise belonged to middle-high frequency band signal.Based on this,firstly,heart sound signal was decomposed by wavelet,and detail wavelet coefficients and approximate wavelet coefficients were selected.Then average Shannon energy envelope and autocorrelation function were used on detail wavelet coefficients and approximate wavelet coefficients respectively,so that wavelet coefficient envelope autocorrelation feature could reflect periodic structural characteristics of heart sound.Next local linear embedding algorithm was used for deminsion reduction.Finally optimized feature set was obtained.For this kind of feature extraction algorithm,three public heart sound databases were used for analysis,and the recognition rate could reach up to 100%.Compared with the first kind of feature extraction algorithm,it had the advantages of short operation time,wide application,which was well compensated the shortrage of first kind of feature extraction algorithm.