Bacteremia prediction models to reduce unnecessary blood cultures: external validation in a large US emergency department.

Bacteremia prediction models can avoid unnecessary blood cultures while missing few bacteremia patients in European cohorts. We aimed to assess the performance of these models in the United States. We externally validated five prediction models in patients ≥ 18 years old evaluated in an academic emergency department (ED) in the United States, between October 2013 and October 2023. Patients needed to have one or more blood cultures and a procalcitonin (PCT) test within 6 h of obtaining blood cultures. The models used basic clinical information such as age and vital signs, with or without the addition of C-reactive protein (CRP), PCT or both. Missing data were imputed. The outcome of the models was bacteremia, defined as at least one true positive blood culture obtained within 48 h of the ED visit. Performance of the models was assessed with the C-statistic for discrimination and calibration plots for calibration. Clinical usefulness was evaluated by the potential reduction in blood cultures, missed bacteremic episodes and decision curve analysis. We included 12,864 patients (55% male, median age 66 years). 1164 out of 12,864 (9.0%) had bacteremia, while 843 (6.6%) patients had contaminated blood cultures. We observed substantial improvement in performance for models that included PCT, compared to those based on clinical information alone or with CRP. The best performing models included PCT with or without clinical information, with C-statistics of 0.80 (95% CI 0.79-0.82 and 0.79-0.81), adequate overall calibration (predicted 10.4%-13.0%, observed 9.0%). Both models were able to reduce ̴30% of blood cultures while missing 5% of bacteremic episodes. We conclude that previously proposed bacteremia prediction models demonstrated robust applicability in a large United States ED population. Models including procalcitonin have promise to safely avoid many blood cultures in the ED. Further validation and impact studies may support the implementation of these models.