Examination of Training Data Expansion for Detection of Abnormal Respiration and Patients

Abstract

Abnormal sounds, termed adventitious sounds, include the lung sound of an individual with pulmonary disease. In this study, we aim to automatically detect abnormal sounds from auscultatory sounds. First, stochastic models are employed to express the acoustic features of normal lung sounds from healthy individuals and abnormal lung sounds from patients. Using this, normal and abnormal lung sounds are classified. However, a low classification rate was obtained because the amount of training data for the stochastic models was small. Although large volumes of training data are necessary for constructing stochastic models with high accuracy, collecting various types of abnormal respiration from a large number of patients is challenging. Therefore, to overcome this limitation, we propose the method to expand the training data for the models. Adding the acoustic features of adventitious sounds to normal respiration and using them as abnormal respiration for training data, significantly increased the classification rate. The results indicate the effectiveness of the proposed method.