Clinical prediction model for bacterial co-infection in hospitalized COVID-19 patients during four waves of the pandemic

The reported estimates of bacterial co-infection in COVID-19 patients are highly variable. We aimed to determine the rates and risk factors of bacterial co-infection and develop a clinical prediction model to support early decision-making on antibiotic use. This is a retrospective cohort study conducted in a tertiary-level academic hospital in Israel between March 2020 and May 2022. All adult patients with severe COVID-19 who had a blood or lower respiratory specimen sent for microbiological analyses within 48 h of admission were included. The primary study endpoint was the prevalence of bacterial co-infection at the time of hospital admission. We created a prediction model using the R XGBoost package. The study cohort included 1,050 patients admitted with severe or critical COVID-19. Sixty-two patients (5.9%) had a microbiologically proven bacterial infection on admission. The variables with the greatest impact on the prediction model were age, comorbidities, functional capacity, and laboratory parameters. The model achieved perfect prediction on the training set (area under the curve = 1.0). When applied to the test dataset, the model achieved 56% and 78% specificity with the area under the receiver operating curve of 0.784. The negative and positive predictive values were 0.975 and 0.105, respectively. Applying the prediction model would have resulted in a 2.5-fold increase in appropriate antibiotic use and an 18% reduction in inappropriate use in patients with severe and critical COVID-19. The use of a clinical prediction model can support decisions to withhold empiric antimicrobial treatment at the time of hospital admission without adversely affecting patient outcomes.IMPORTANCEEstimates of bacterial coinfection in COVID-19 patients are highly variable and depend on many factors. Patients with severe or critical COVID-19 requiring intensive care unit admission have the highest risk of infection-related complications and death. Thus, the study of the incidence and risk factors for bacterial coinfection in this population is of special interest and may help guide empiric antibiotic therapy and avoid unnecessary antimicrobial treatment. The prediction model based on clinical criteria and simple laboratory tests may be a useful tool to predict bacterial co-infection in patients hospitalized with severe COVID-19. Estimates of bacterial coinfection in COVID-19 patients are highly variable and depend on many factors. Patients with severe or critical COVID-19 requiring intensive care unit admission have the highest risk of infection-related complications and death. Thus, the study of the incidence and risk factors for bacterial coinfection in this population is of special interest and may help guide empiric antibiotic therapy and avoid unnecessary antimicrobial treatment. The prediction model based on clinical criteria and simple laboratory tests may be a useful tool to predict bacterial co-infection in patients hospitalized with severe COVID-19.