A transfer learning-based deep learning approach for automated COVID-19 diagnosis with audio data

Abstract

The COVID-19 pandemic has caused millions of deaths and changed daily life globally. Countries have declared a half or full lockdown to prevent the spread of COVID-19. According to medical doctors, as many people as possible should be tested to identify their status, and corresponding actions then should be taken for COVID-19 positive cases. Despite the clear necessity of these medical tests, many countries are still struggling to acquire them. This fact clearly indicates the necessity of a large-scale, cheap, fast, and accurate alternative prescreening tool that can be used for the diagnosis of COVID-19 while waiting for the medical tests. To this end, a novel end-to-end transfer learning-based deep learning approach that uses only a given cough sound for the diagnosis of COVID-19 was proposed in this study. The proposed models employed various pretrained deep neural networks, namely, VGG19, ResNet50V2, DenseNet121, and MobileNet, via the transfer learning technique. Then, these models were evaluated on a gold standard dataset, namely, Cambridge data. According to the experimental result, the proposed model, which employed the MobileNet via the transfer learning technique, provided the best accuracy, 86.42%, and outperformed the state-of-the-art. Thus, the proposed model has the potential to provide automated COVID-19 diagnosis in an easily applicable and fast yet accurate way.