Detection of abnormal lung sounds considering spectral and temporal features of heart sounds

In this paper, we propose a robust classification method for lung sounds contaminated with heart sounds in order to distinguish between healthy subjects and abnormal patients with pulmonary emphysema. We previously developed a classification procedure based on a maximum-likelihood approach by using hidden Markov models (HMMs). However, contaminated heart sounds caused difficulties in achieving a highly accurate classification, because it was difficult to generate HMMs that distinguished between adventitious sounds and heart sounds with high accuracy, by using power and spectral acoustic features only. To address this problem, we propose a classification technique that is based on the use of spectral features and temporal features related to heart sounds: distributions of durations and time intervals of heart (S1) sounds. A validity score for detected adventitious sounds and heart sounds in the classification process is designed by considering the distribution of time intervals of heart sounds and differences in the durations between the adventitious sounds and the heart sounds. In the proposed method, the total likelihood of each respiratory sound is obtained by summing the spectral likelihood derived from the HMMs and the validity score. In the classification of healthy subjects and patients using 94 lung sound samples from 94 subjects, the proposed method achieved a higher classification rate (90%) than the baseline method (84%) using only the spectral features, thus demonstrating the superiority of the proposed method.