Training Data Expansion for Classification between Normal and Abnormal Lung Sounds

Abstract

In this paper, we investigate the effectiveness of training data expansion methods to distinguish between normal and abnormal lung sounds. Acoustic characteristics of lung sounds vary according to auscultation points. In conventional classification methods, acoustic models were usually trained using only lung sounds recorded at the same auscultation points to that of evaluation data. This results in a small amount of training data and, thus, hinders the achievement of a high classification rate. To overcome this problem, we performed training data expansion by selecting the lung sounds, which are expected to be useful for generating acoustic models with higher classification performance, among sound samples recorded at other auscultation points. We investigated the two types of selection approach: selection based on the similarity of acoustic features in sound samples and selection based on the confidence measure represented by the difference between the acoustic likelihood for a normal or abnormal respiratory candidate. Our experiments showed that both selection types have the potential to increase the classification performance between normal and abnormal lung sounds, as well as the classification performance between healthy and unhealthy subjects.