Application and Performance Improvement of Transfer Learning on ICBHI Lung Sound Dataset

Chronic respiratory diseases are one of the leading causes of morbidity and mortality worldwide. How to prevent the disease or to diagnose and treat it effectively in the early stage has always been a focused medical research area. In this paper, a neural network that was pre-trained based on a large audio event dataset called AudioSet is transferred and applied in the training and testing of the Respiratory Sound database ICBHI; in addition, various methods are used in data preprocessing, neural network configuration and post-processing to improve the performance of the transfer learning model. The final model can not only converge quickly, but also use the accuracy calculation method provided by ICBHI Challenge to reach 81.1 \(\%\) in the four classification tasks containing normal, crackle, wheeze and both respiratory sounds, which is superior to the previous methods. This paper also analyzes the unbalanced distribution of the respiratory cycle dataset based on demographic data on the binary classification task (normal and abnormal). The binary classification model scored 85.5 \(\%\) and 81.1 \(\%\) on the female test group and the male test group, respectively. To address the above differences due to the unbalanced dataset, we used a restricted mixup approach to successfully reduce the difference between the male and female test groups to 0.82 \(\%\) .