Self-supervised learning for intelligent disease diagnosis using audio signals: beyond copd to a spectrum of diseases

Given the widespread prevalence and significant patient base of COPD (Chronic Obstructive Pulmonary Disease), the development of simple and rapid diagnostic methods has emerged as a key research focus. Through pathological studies, the medical community has identified the potential of cough sounds for diagnosing COPD, sparking interest in leveraging deep learning to analyze various disease-related sounds, including those associated with COVID-19 and cardiac conditions, etc. Yet, research specifically targeting COPD remains scarce, primarily due to two challenges: traditional models trained on small medical datasets often fall short of expectations due to stringent data privacy and collection requirements in healthcare; and the scarcity of publicly accessible COPD datasets, particularly those that could obviate the need for medical equipment. Addressing these challenges, our paper introduces a novel dataset of smartphone-recorded cough sounds, termed the CC (COPD-Cough) dataset. It comprises 221 recordings from COPD patients and 632 from healthy individuals, marking the first dataset explicitly curated for COPD cough sound analysis. The dataset, endorsed by clinical professionals and collected independently of medical devices, promises to propel advancements in straightforward COPD diagnostics. Furthermore, we propose a self-supervised learning model enhanced by unique data augmentation techniques and an efficient sound feature extractor, demonstrating superior performance across three distinct disease datasets and achieving state-of-the-art results. Comprehensive ablation studies affirm our model’s efficacy, while sensitivity analyses optimize its applicability to various tasks. For further engagement, the framework’s source code and dataset are available at https://github.com/auto-chao/COPD_Diagnosis and https://zenodo.org/records/10209837, respectively.