Investigation of noise effect on lung sound recognition

Abstract

More studies are needed to evaluate the effect of the experimental results in noisy environments if lung sound recognition system is to be used as a reliable bedside monitoring equipment. In this study, three feature representations: autoregressive coefficients, Mel-frequency cepstral coefficients, and bispectrum diagonal slices, were utilized to characterize lung sound signal. In order to compare the performance of the three feature types under various noise conditions, the white Gaussian and two real noises (babble and car noises) with various SNR levels were added to each lung sound signal in the test database. The dynamic timing warping was selected as classifier to discriminate the lung sounds to one of the three categories: normal, wheeze, or crackle. Our experimental results showed that the bispectrum diagonal slices was more immune to noise interference in lung sound recognition, but the Mel-frequency cepstral coefficients was more vulnerable to noise disturbance.