New approaches for spectro-temporal feature extraction with applications to respiratory sound classification

Auscultation based diagnosis of pulmonary disorders relies on the presence of adventitious sounds. In this paper, we propose a new set of features based on temporal characteristics of filtered narrowband signal to classify respiratory sounds (RSs) into normal and continuous adventitious types. RS signals are first decomposed in the time–frequency domain and features are extracted over selected frequency bins containing distinct signal characteristics based on auto-regressive averaging, the recursively measured instantaneous kurtosis, and the sample entropy histograms distortion. The presented features are compared with existing features using a modified clustering index with different distance metrics. Mean classification accuracies of 97.7% and 98.8% for inspiratory and expiratory segments respectively have been achieved using Support Vector Machine on real recordings.