1683: Development of a Bayesian Risk Assessment of Infection (Brain) Model for Hospital Admissions

Introduction: Inflammatory mediators are often used to assess the risk of bacterial infection. Frequentist analysis does not allow the incorporation of clinical judgment in the decision-making. This study was conducted to describe clinical phenotypes based on C-reactive Protein (CRP) and Procalcitonin (PCT) and to develop a Bayesian prediction model for the posterior probability of bacterial infection in patients presenting to the emergency department (ED). Methods: Retrospective cohort study (15 hospitals;2019-2023). Patients admitted from the ED who had CRP, procalcitonin, and bacterial cultures sent within the first 24 hours were included. Patients were classified into four groups (both CRP and PCT normal group A, both abnormal group B, only PCT abnormal group C, only CRP abnormal group D). The Bayesian regression model was developed under three informative clinical priors (0.3, 0.5, and 0.7) and included age, CRP, PCT, fever, white blood cell count, ESR, ferritin, and viral positivity. A prediction calculator was also developed using R-shiny for bedside clinical application. Model validation with posterior predictive check and LOOCV. Results: 10,397 patients (median age 65; 909 < 18 years) were included. The culture positivity rate was 27.5%, and mortality rate was 11.2%. There was a significant difference in culture positivity and mortality between the groups, with the highest culture positivity and mortality in Group B(35% & 17%), followed by Group D (24% & 11%). The Bayesian regression intercept was negative and varied across models. The models showed posterior probabilities of bacterial infection of 25%, 17.5%, and 10.6%, respectively. The individual factors showed stable coefficients across models, PCT was the strongest predictor (1 log unit PCT associated with a 45% increase in probability of bacterial infection). Ferritin and viral positivity associated with a lower probability. AUROC of the three models was 0.64, and AUPRC was 0.43. Conclusions: Inflammatory mediators are often discordantly elevated in patients with suspicion of bacterial infection. The Bayesian calculator (BRAIN) allows users to calculate the posterior probability of bacterial infection with user input of clinical parameters, and app returns individualized risk estimates under different levels of clinical suspicion.