RESUMO
OBJECTIVE: SARS-CoV-2 disease (COVID-19) has become a pandemic disease, determining a public health emergency. The use of artificial intelligence in identifying easily available biomarkers capable of predicting the risk for severe disease may be helpful in guiding clinical decisions. The aim of the study was to investigate the ability of interleukin (IL)-6, troponin I, and D-dimer to identify patients with COVID-19 at risk for intensive care unit (ICU)-admission and death by using a machine-learning predictive model. PATIENTS AND METHODS: Data on demographic characteristics, underlying comorbidities, symptoms, physical and radiological findings, and laboratory tests have been retrospectively collected from electronic medical records of patients admitted to Policlinico A. Gemelli Foundation from March 1, 2020, to September 15, 2020, by using artificial intelligence techniques. RESULTS: From an initial cohort of 425 patients, 146 met the inclusion criteria and were enrolled in the study. The in-hospital mortality rate was 15%, and the ICU admission rate was 41%. Patients who died had higher troponin I (p-value<0.01) and IL-6 values (p-value=0.04), compared to those who survived. Patients admitted to ICU had higher levels of troponin I (p-value<0.01) and IL-6 (p-value<0.01), compared to those not admitted to ICU. Threshold values to predict in-hospital mortality and ICU admission have been identified. IL-6 levels higher than 15.133 ng/L have been associated with a 22.91% risk of in-hospital mortality, and IL-6 levels higher than 25.65 ng/L have been associated with a 56.16% risk of ICU admission. Troponin I levels higher than 12 ng/L have been associated with a 26.76% risk of in-hospital mortality and troponin I levels higher than 12 ng/L have been associated with a 52.11% risk of ICU admission. CONCLUSIONS: Levels of IL-6 and troponin I are associated with poor COVID-19 outcomes. Cut-off values capable of predicting in-hospital mortality and ICU admission have been identified. Building a predictive model using a machine-learning approach may be helpful in supporting clinical decisions in a more precise and personalized way.
