An Effective Respiratory Disease Diagnosis System Using Adaptive Oversampling and Dynamic Audio Concatenation

Doctors’ subjectivity in respiratory disease detection through lung auscultation needs to be overcome. In this study, we propose a deep learning architecture as a recognition system. We introduce the ResNet-18 model, coupled with adaptive oversampling to address the data imbalance issue in the SPRSound database 2022, and combined with dynamic audio concatenation for feature enhancement. Ultimately, our system achieves the highest performance with a score of 0.820 in the seven-class classification task.