Detecção de indivíduos infectados com Covid-19 através de sons de tosse forçada

Abstract

The new coronavirus (COVID-19) is an infectious disease declared a pandemic in 2020 by the World Health Organization (WHO). Through a lot of cooperation and the effort of scientists around the world, several vaccines were created. However, even if a large part of the world’s population is vaccinated, there is no guarantee that the virus will ever disappear. Therefore, non-invasive methods, with low cost, and capable of generating results in real-time are important to detect infected individuals and enable an adequate treatment quicker, in addition to preventing the spread of the virus. In the literature it is possible to find computational methods capable of distinguishing a healthy person from an infected with COVID-19 with high accuracy, using data sets of forced cough, collected online from individuals around the world. A great existing challenge is in the unbalance of these data, considering that there are more samples of healthy individuals than contaminated ones. In this work, we proposed changes in a Deep Neural Networks model, available in the literature, as well as in its training. In addition, we carried out studies on the data augmentation of these audios. The results show that the adapted model is simpler (having fewer parameters) and manages to better generalize the prediction of infected individuals, presenting an average Area Under the Curve ROC (AUC) of 0.885 and a confidence interval (0.881 - 0.888), against 0.771 and (0.752 - 0.783) of the original one.