Abstract:
The objective of this study is to perform two complementary analyses of pulmonary crackles, i.e. modeling and clustering, in order to interpret crackles in timefrequency domain and also determine the optimal number of crackle types and their characteristics using the modeling parameters. Since the crackles are superimposed on background vesicular sounds, a preprocessing method for the elimination of vesicular sounds from crackle waveform is also proposed for achieving accurate parameterization. The proposed modeling method, i.e. the wavelet network modeling, interprets the transient structure of crackles in the time-frequency space with a small number of components using the time-localization property of wavelets. In modeling analysis, complex Morlet wavelets are selected as transfer functions in the hidden nodes due to both their similarity with the crackle waveforms and their exibility in the modeling process. Clustering analysis of crackles probe the discrepancies found among the studies related with the crackle types and their corresponding characteristics. Since, in these studies, crackles are classi ed according to the auditory perception of the observers, there are inconsistencies found in the labeling of the same crackle. To overcome the inherent subjectivity, the crackles are classi ed in an unsupervised method using the EM clustering analysis. In this method, it is assumed that the crackle data can be interpreted with the multivariate Gaussian mixture model and, therefore, crackle clusters distribute normally in the feature spaces. The results strongly suggest the existence of a third crackle type, medium, in addition to the commonly used two types, i.e. ne and coarse. Moreover, the extracted characteristics of crackle types o er additional features for the computerized crackle-based analysis of pulmonary disorders.|Keywords: Lung Sounds, Pulmonary Crackles, Crackle Types, Wavelet Networks, Signal Modeling, EM Clustering, Vesicular Sound Elimination
