Modelo preditivo para prognóstico de pacientes com COVID-19

Abstract

With the criticality of the lack of beds in the intensive care unit during the COVID-19 pandemic, policies to determine who has access to beds were implemented in Brazilian states. Aiming to mitigate the subjectivity of these policies, we propose the development and internal validation of a predictor for death risk classification of patients with COVID19 in Rio Grande do Sul. We use the TRIPOD guideline to explain the development of the random forest-based machine learning model. The dataset has 604,389 records and includes patients treated in Rio Grande do Sul who were reported through the Painel Coronavírus RS, in the period from January 1 to June 8, 2021. The outcome variable is evolution, which informs whether the patient was recovered or died from COVID19. In total, fourteen characteristics were listed as predictors, which were demographic (gender and age group) and clinical (symptoms and comorbidities). The derivation dataset has 408,959 records, being 3.18% (n=13,005) deaths, and the validation set has 175,269 records, being 5,574 deaths. In the test set, the model rated the chance of death with an AUC-ROC score of 0.981 (95% confidence interval 0.981 to 0.982) and AUC-ROC of 0.970 in the validation set. The classifier makes it possible to stratify with good precision the risk of death of patients with COVID-19 and can contribute to the reduction of subjectivity in decision-making in the hospital environment. However, we point out that this is just an additional tool, and all ethical and legal aspects must be considered in decision-making. Key-words: Machine Learning. Patient Priorization. Risk Stratification. COVID-19 Prognosis.