Observation, parametric modelling and classification of respiratory sounds

Abstract

Auscultation is a widely used method in the diagnosis ofpulmonary diseases and in the analysis of respiratory sounds. Thecharacteristics of respiratory sounds show differences in pathological casesfrom normal cases. The object of this study is, to observe the characteristicsof respiratory sounds in both cases, to analyse them in time and frequencydomain and to distinguish a normal case from a pathological case. Toachieve mentioned goals, respiratory sounds heard over the chest waHfrom the specific locations were recorded. The flow signal was alsorecorded by a flowmeter to synchronize on the inspiration and expirationphases, because the characteristics of respiratory sounds may change fromphase to phase. An AR modeling was applied to obtain a parametricrepresentation of the sounds. The analysis of respiratory sounds wasperformed after they were distinguished to inspiration and expirationphases. Mahalanobis distance measure, and minimum distance classificationmethod is used to classify respiratory sounds into appropriate classes.Experiments showed that the suggested classifier can distinguish thenormal case from a pathological case if and only if a large database of lungsounds is available. The classification method was also compared withItakura distance measure and k-nearest neighbor classification methodwhich was performed in a previous study. The abrupt changes (crackles) inthe respiratory sound waveforms of pathological cases were observed anda new method is suggested to detect them because they have a specialimportance in the diagnosis of some pulmonary diseases.