Aplicações de aprendizado de máquina para previsão de mortalidade da Covid-19 usando o banco de dados nacional DATASUS

Abstract

An important recent health issue for the nations of the world has been the coronavirus pandemic. COVID-19 is associated with di erent clinical manifestations accompanied by a high mortality rate. In Brazil we have a national repository of clinical data - DATASUS - which, among the data on various pathologies, are those associated with COVID-19. Predicting mortality through outcome predictors may be a crucial tool in the care of patients a ected by the disease. Multivariate methods and machine learning can be used to develop prediction models mortality rate in order to assist health professionals. In this work we use the Perceptron, MLP and Random Forest models to obtain these results in databases segmented by days of hospitalization. Among these methods, the one that obtained the best result was the Random Forest, with an average accuracy of 82%. In addition to implementing these techniques, we also suggest a new method of lling missing values that we use in the DATASUS database itself, besides comparing it to other techniques on benchmark databases.