Pediatric Respiratory Sound Classification Using a Dual Input Deep Learning Architecture

Abstract

Patients with respiratory conditions typically exhibit adventitious respiratory sounds (ARS), such as wheezes and crackles. In recent years, computerized methods for analyzing respiratory function, namely ARS, have gained increased attention within the scientific community. Such methods primarily aim to facilitate diagnosing and monitoring patients suffering from respiratory diseases. In this work, we propose a deep learning model for the automatic classification of respiratory sounds within the proposed tasks of the "IEEE BioCAS 2023 Grand Challenge on Respiratory Sound Classification". The model was based on a dual input convolutional deep learning architecture, using the raw audio signal and the short-time Fourier transform (STFT) spectrogram as inputs. Our model obtained a challenge total score of 0.590 (Task 1-1: 0.756; Task 1-2: 0.467; Task 2-1: 0.658; Task 2-2: 0.458).