Classification between normal and abnormal lung sounds using unsupervised subject-adaptation

In this paper, we propose an unsupervised subject-adaptation method to distinguish between normal and abnormal lung sounds. Lung sounds have varied subject-dependent acoustic characteristics. In conventional classification methods using subject-independent acoustic models, this diversity hinders the achievement of a high classification rate. To overcome this problem, we performed unsupervised subject-adaptation of acoustic lung-sound models by exclusively employing respiration periods that could be confidently considered normal/abnormal in test respiration input from an unknown subject. In our method, these confident periods were detected based on the difference between the acoustic likelihood for a normal respiratory candidate and that for an abnormal candidate. The proposed adaptation method achieved a higher classification performance of 83.7% between normal and abnormal respiration in comparison with the baseline method that did not use adaptation, which achieved a performance of 82.7%. Our method for classifying healthy subjects and patients with pulmonary emphysema achieved a higher classification rate of 84.3% relative to the baseline (83.5%).