A model to predict SARS-CoV-2 infection based on the first three-month surveillance data in Brazil

Abstract

Objective: COVID-19 diagnosis is a critical problem, mainly due to the lack or delay in the test results. We aimed to obtain a model to predict SARS-CoV-2 infection in suspected patients reported to the Brazilian surveillance system. Methods: We analysed suspected patients reported to the National Surveillance System that corresponded to the following case definition: patients with respiratory symptoms and fever, who travelled to regions with local or community transmission or who had close contact with a suspected or confirmed case. Based on variables routinely collected, we obtained a multiple model using logistic regression. The area under the receiver operating characteristic curve (AUC) and accuracy indicators were used for validation. Results: We described 1468 COVID-19 cases (confirmed by RT-PCR) and 4271 patients with other illnesses. With a data subset including 80% of patients from Sao Paulo (SP) and Rio Janeiro (RJ), we obtained a function which reached an AUC of 95.54% (95% CI: 94.41–96.67%) for the diagnosis of COVID-19 and accuracy of 90.1% (sensitivity 87.62% and specificity 92.02%). In a validation dataset including the other 20% of patients from SP and RJ, this model exhibited an AUC of 95.01% (92.51–97.5%) and accuracy of 89.47% (sensitivity 87.32% and specificity 91.36%). Conclusion: We obtained a model suitable for the clinical diagnosis of COVID-19 based on routinely collected surveillance data. Applications of this tool include early identification for specific treatment and isolation, rational use of laboratory tests, and input for modelling epidemiological trends.