Detection of Adventitious Respiratory Sounds based on Convolutional Neural Network

Abstract

Nowadays, the respiratory disease has become one of the most dangerous diseases that threaten human health, especially in the developing countries. The early diagnosis of respiratory disease gives patients the opportunity to receive proper treatment in time, and hence artificial intelligent (AI) auscultation using electronic stethoscope may play a promising role here. The core idea of AI auscultation of respiratory disease is to detect or recognize two kinds of adventitious respiratory sounds related to respiratory diseases: wheeze and crackle. Constrained by the number of available data, in the traditional methods, subjectively defined features were extracted and used to detect these adventitious respiratory sounds. However, to make the detection results robust, the features had better to be learned automatically from the data, which can be realized by applying deep learning in a big data. In this paper, the convolutional neural network (CNN) is exploited to detect adventitious sounds. The data used in this study consists of two parts: the public database provided by the International Conference on Biomedical and Health Informatics (ICBHI) involving 126 subjects and our recorded pediatric auscultation data including 222 subjects. The detection performance of employed CNN is evaluated using ICBHI database, our pediatric auscultation database as well as the combination of them.