Cough sound analysis using Deep Learning methods for COVID-19 diagnosis

Abstract

COVID-19 (COronaVIrus Disease of 2019), caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), has been challenging humanity for the past one and a half year. It can cause severe illness and dysfunction in multiple human organs, with many patients finally passing away. Although vaccines have been released and are widely used around the globe, the essential amount of immunity in order for the COVID-19 transmission between people to terminate, has not been reached yet. Ever since the beginning of this pandemic, the frequent testing of large portions of the population played a determinant role in the containment of the spread. However, the two widely used testing methods, Nucleic Acid Amplification Tests (NAATs) and antigen tests are time and fund consuming, obstructing the screening process of large groups of people. Moreover, the transmission of possible cases to health structures involves the risk of contaminating both the personnel and the rest of the patients. The current thesis examines a different screening method, which is both time and cost efficient and does not require the transportation of individuals to health facilities. The method used, leverages the success of Machine Learning and especially Convolutional Neural Networks (CNNs) for the detection of COVID-19 through cough samples recorded by the mobile phone of the user or a web application. The cough samples collected are converted to images and fed into a CNN architecture which is trained to classify them between COVID-19 and non-COVID-19. One of the main challenges of COVID-19 cough classification lies in the fact that cough is a symptom of multiple non-COVID-19 related medical conditions. Moreover, the high imbalance of the available datasets, with the COVID-19 samples being significantly less than the non-COVID-19 samples and the fact that the datasets are crowd-sourced, are two important factors making the current task demanding. That is due to the entailed difficulty of using non clean data, with a ground truth based on the declarations of the users. More specifically, the samples provided by each user may contain sounds of low quality, while the validity of the information relative to the user being positive or negative to COVID- 19 cannot be confirmed. To that end, different methods, architectures and datasets were examined. A 5-fold cross validation approach was used examining different combinations of datasets and architectures. In order to deal with the imbalanced nature of the data, an ensemble learning method was implemented. Since Deep Learning architectures are data "hungry", training them with multiple datasets was also examined, providing the highest classification results with an accuracy of 71.60%. The obtained results certify the ability of detecting COVID-19 infection through cough sounds, but more importantly the ability of using Machine Learning for the diagnosis of respiratory diseases. This could play a determining role in the quicker containment of future pandemics.