Avaliação do aprendizado ativo na predição de óbitos por Covid-19 com machine learning

Abstract

The SARS-CoV-2 virus, the causative agent of COVID-19, triggered the most recent pandemic to affect humanity. This disease, in its most serious phase, attacks the respiratory system and can lead to death. Therefore, the ability to classify COVID-19 cases as mild or severe is of utmost importance in treatment. The use of traditional Machine Learning (ML) methods can be applied to predict severe cases, however, these methods face challenges in dealing with the dynamics of the disease, resulting in errors, especially in critical age groups. It is believed that Active Learning (AL) can adapt to the problem, allowing a more effective adaptation to the dynamics of the disease, training the model with only the most relevant data, as opposed to the training approach used by traditional ML methods. In this way, models will be created with AL to try to improve the prediction of deaths from COVID-19. To evaluate the models, the AUC (Area Under the ROC Curve) and the false negative rate (FNR) were used for age groups and in general. The data used is that of Severe Acute Respiratory Syndrome (SRAG) including cases of COVID-19 made available by the Unified Health System (SUS) through the OpenDataSus website. The best model using AL stood out in the age groups from 61 to 70 to 81 to 90 years old, which generally obtained an FNR of 0.4852 in contrast to traditional ML methods that obtained an overall FNR of 0.528.