Bacteraemia is associated with poor outcomes, but the diagnostic gold standard, peripheral blood culture, can take up to 24 hours to inform management. Single predictors and existing sepsis risk scores perform poorly for early identification. To address this gap, the authors developed a multivariable logistic regression model aimed at predicting bacteraemia in adults presenting to emergency departments using data routinely available within hours of hospital attendance.
Model development used a retrospective cohort of 33,874 hospital encounters at University College London Hospitals (UCLH) spanning 2019 to 2024. Temporal validation was performed on held-out UCLH data from 2024. External validation was conducted in a large independent dataset comprising 53,669 hospital encounters from the Infections in Oxfordshire Research Database (IORD). Reported bacteraemia prevalence was 5.2% in the UCLH development cohort and 8.9% in the IORD external validation cohort.
Twenty predictors were retained in the final model. The predictors spanned patient demographics, comorbidities, vital signs, and routine blood tests. Selection of predictors followed backward selection guided by the Akaike information criterion (AIC) from a set of predefined candidate variables that are typically available within hours of hospital attendance. Continuous predictors were modelled using restricted cubic splines to capture potential non-linear associations.
The modelling approach was logistic regression with backwards AIC-based selection. Missing data were handled using multiple imputation. The authors assessed model stability with internal–external cross validation and prediction instability analysis prior to temporal and external validation. Where relevant, model performance statistics are reported with 95% confidence intervals.
Discrimination in the development assessments was stable across time periods with a pooled c-statistic of 0.82 (95% CI 0.81 to 0.84). Temporal validation in held-out 2024 UCLH data produced a c-statistic of 0.83 (95% CI 0.79 to 0.87). External validation in the IORD cohort yielded a c-statistic of 0.83 (95% CI 0.82 to 0.83). Calibration in external validation was excellent: calibration slope 1.08 (95% CI 1.05 to 1.11) and calibration-in-the-large 0.01 (95% CI −0.02 to 0.04). These metrics indicate consistent discrimination and accurate absolute risk estimation in an independent population.
Internal validation employed internal–external cross validation and prediction instability analysis to evaluate stability across development periods. Temporal validation used held-out UCLH encounters from 2024. External validation was performed in the IORD dataset of 53,669 encounters. Performance measures were maintained across these validations, supporting the model’s generalisability across different UK hospital populations and time periods.
The model outperformed single predictors, established sepsis risk scores, and a reconstructed comparator model drawn from previous work. Decision curve analysis showed superior net benefit compared with these alternatives, suggesting the model offers greater clinical utility for decision-making around suspected bacteraemia when compared with commonly used single measures or risk scores.
Performance was consistent across subgroups defined by age, sex, ethnicity, and socioeconomic status. However, model accuracy degraded when blood cultures were sampled more than six hours after attendance, and discrimination varied by likely site of infection. These findings imply that timeliness of sampling and infection focus influence predictive accuracy and should be considered in implementation.
The model uses routinely collected variables available within hours of emergency department attendance to provide an early, interpretable probability of bacteraemia. Given stable discrimination and excellent external calibration, the tool could support early clinical decision-making—such as targeting antimicrobial therapy, prioritising diagnostics, or guiding admission and observation—once implementation thresholds and action strategies are defined. The authors emphasise the model’s generalisability but note that prospective implementation work is needed.
This work is reported as a preprint and has not been peer reviewed. The authors note variation in performance by timing of culture sampling and infection site, and they indicate that determining optimal implementation thresholds remains necessary. Further work was recommended to translate predicted probabilities into actionable clinical pathways and to evaluate prospective impact on patient outcomes and clinical workflow.