Hypergradient Descent Based Multi-Task Learning on Auscultation Point Guided Respiratory Sound Classification

Respiratory diseases, highlighted by the COVID-19 pandemic, are now one of the leading causes of global mortality. Auscultation is an essential diagnostic tool for these conditions. However, its precision is dependent on the proficiency of healthcare providers, which can be a limitation in areas with scarce medical resources. Recently, the advent of sophisticated electronic stethoscopes and the application of deep learning have propelled the development of respiratory sound classification technologies to aid in clinical diagnosis. Yet, the variability of clinical auscultation environments has hindered the performance of these technologies. In this study, we propose a novel approach to enhance the model's performance through multi-task learning (MTL), simultaneously capturing the acoustic features of different respiratory sound types and their corresponding auscultation points. We've observed that normal respiratory sounds differ based on auscultation points, and the scarcity of data can confound models in discerning whether feature variations are point-related or indicative of pathology. By integrating both learning tasks, we overcome this challenge. To reinforce model robustness, we employ hyper-gradient descent (HD) to balance the task weights. Our work achieves a score of 62. 98% in the ICBHI data set, which is the arithmetic mean of Specificity and Sensitivity, surpassing the baseline by 3.43 % and demonstrating state-of-the-art performance. We believe that our findings can serve as inspiration and be integrated with other works employing data augmentation techniques to further enhance the performance and generalizability of models in clinical settings at large.