Design and development of a Bayesian risk assessment model for bacterial infection (BRAIN) in patients admitted to...

C-reactive Protein (CRP) and Procalcitonin (PCT) are commonly used in conjunction with clinical judgment to assess the risk of bacterial infection. Traditional frequentist methods do not allow the incorporation of clinical suspicion into risk estimation. This study aimed to describe clinical phenotypes based on CRP and PCT levels and develop a Bayesian model to estimate the posterior probability of bacterial infection in emergency department (ED) patients.

Retrospective study across 15 hospitals (2019-2023), including patients admitted from the ED with CRP, PCT, and bacterial cultures obtained within 24 h. Patients categorized into four groups: both markers normal (A), both abnormal (B), only PCT abnormal (C), and only CRP abnormal (D). Bayesian logistic regression models developed using priors of 0.3, 0.5, and 0.7 to reflect varying levels of clinical suspicion. Predictors included age, CRP, PCT, fever, white blood cell count, ESR, ferritin, and viral positivity.

Among 10,397 patients (median age, 65; 909 < 18 years), 27.5% had positive cultures, with an 11.2% mortality rate. Culture positivity and mortality were highest in group B (35%), followed by D (24%). Posterior probabilities of infection under high, moderate, and low suspicion priors were 25%, 17.5%, and 10.6%, respectively. PCT was the strongest predictor, with a one log-unit increase associated with a 45% rise in infection probability. AUROC was 0.64; AUPRC 0.43. A R Shiny calculator (BRAIN) was created for bedside application.

A Bayesian model incorporating inflammatory markers and clinical judgment provides individualized estimates of bacterial infection risk at the bedside.