Wheeze detection in respiratory sounds via statistical signal modeling

Abstract

The aim of this study is detection of wheeze and non-wheeze epochs within respiratory sound signals acquired from patients with asthma and COPD. Since a wheeze signal, having a sinusoidal waveform, has a different behavior in time-frequency domain from that of a non-wheeze signal, the features selected for detection are Renyi entropy, f50/f90 ratio and mean-crossing irregularity. Upon calculation of these features for each wheeze and non-wheeze portion, two approaches are proposed. In the first approach (multi-dimensional approach) the whole data scattered as two classes in three dimensional feature space are assumed to be Gaussian distributed. Then, a decision rule is applied for two different Gaussian random vectors. In the second approach (one-dimensional approach) the three dimensional data are projected onto the single dimensional space that separates the two classes best by using Fisher Discriminant Analysis (FDA). These one-dimensional data are also assumed to be Gaussian distributed. Then, a decision rule is applied for two different Gaussian variables. Finally when the total number of false positives and false negatives are considered, the minimum value of probability of error over the data set is found to be 0.045 for both approaches and the Chernoff upper bounds for this error are found as 0.09 and 0.10 for multi and one-dimensional approaches, respectively.