Related Experiment Videos
Development and validation of a bacteremia prediction score for adult emergency department patients using hybrid
Minha Kim1,2, Min Joon Seo3, Jong Eun Park1
1Department of Emergency Medicine, Samsung Medical Center, Sungkyunkwan University School of Medicine, 115 Irwon-ro, Gangnam-gu, Seoul, 06355, Republic of Korea.
Abstract:
Bacteremia requires early detection and appropriate treatment. High rates of false-positives in blood cultures in emergency departments (EDs) have led to the need for accurate predictive tools. This study developed and validated a bacteremia prediction score to improve patient selection for blood cultures. A multicenter retrospective cohort study used data from a tertiary academic hospital in Republic of Korea (2014–2019) to develop and validate the model, with external validation at two additional hospitals (Hospital A and Hospital B) (2022). The AutoScore method, integrating machine learning and logistic regression, generated the score using initial vital signs and laboratory data. Model performance was assessed via the area under the receiver operating characteristic (AUROC) curve. Of 24,856 patients, 17,373 formed the development cohort, 2491 the validation cohort, and 4992 the test cohort. Bacteremia incidence was 11.39%. The score, incorporating eight variables (procalcitonin, lactate, white blood cell count, c-reactive protein, blood urea nitrogen, body temperature, platelet count, and total bilirubin), achieved an AUROC of 0.825 (95% confidence interval [CI], 0.808–0.842), outperforming procalcitonin (0.808; 95% CI, 0.790–0.826), SOFA score (0.718; 95% CI, 0.696–0.741), MEWS score (0.630; 95% CI, 0.606–0.654) and SIRS score (0.583; 95% CI, 0.558–0.608). External validation yielded AUROCs of 0.796 (95% CI, 0.773–0.818) at Hospital A and 0.813 (95% CI, 0.778–0.848) at Hospital B. Risk stratification categorized patients into low-, intermediate-, and high-risk groups with bacteremia rates of 2.6–3.2%, 5.2–9.3%, and 23.7–30.2%. A bacteremia prediction score was developed and validated using AutoScore, integrating machine learning and logistic regression. The model, based on eight clinical variables, stratified patients by risk and showed favorable predictive performance compared with conventional markers.
