Effect of Auscultation Hindering Noises on Detection of Adventitious Respiratory Sounds Using Pretrained Audio Neural Nets: A Comprehensive Study

Respiratory disorders cause the death of around 8 million individuals globally. These disorders can be diagnosed by utilizing different clinical modalities, such as spirometry, FeNO test, and peak-flow measurements. However, chest auscultation-based respiratory sound (RS) examination with stethoscopes remains one of the most important clinical tools for identifying such disorders, since different adventitious respiratory sound cycles (ARSCs) correlate with various structural flaws of the lung. Therefore, an efficient identification of the ARSs using deep learning (DL)-based computerized algorithms, will help in the early detection of respiratory disorders. However, the RSs are highly susceptible to different types of interferences when collected in a real-world clinical setting, which might impair the precision of the ARSC classification algorithms and increase the chances of misdiagnosis. This work aims to investigate the effect of various auscultation-hindering noises such as heart sounds (HS), and background ambient noises on the automated classification of ARSCs using the DL algorithms. In this paper, we have performed the ARSC classification experiments under various noisy circumstances by feeding the mel spectrogram representations into different pre-trained audio neural networks (PANNs). Experimental results show that by using the proposed approach, we have achieved an ICBHI score (Mean [95% confidence interval]) of 81.14 ([80.18-82.10]) %, 77.08 ([74.36-79.81]) %, and 70.69 ([69.10-72.28]) %, while classifying the clean or reference RSs, RSs corrupted with HS, and hospital ambient noises, respectively.