Development and validation of a clinical prediction rule for candidemia in hospitalized patients with severe sepsis and septic shock.
J Crit Care. 2015 Aug;30(4):715-20
Authors: Guillamet CV, Vazquez R, Micek ST, Ursu O, Kollef M
OBJECTIVE: To develop and internally validate a prediction rule for the presence of candidemia in patients with severe sepsis and septic shock (candidemia rule) that will fill the gap left by previous rules. To compare the accuracy of the available Candida prediction models.
DESIGN: Retrospective cohort study.
SETTING: Barnes-Jewish Hospital, St. Louis, Missouri.
PATIENTS/SUBJECTS: Two thousand five hundred ninety-seven consecutive patients with a positive blood culture and severe sepsis or septic shock.
INTERVENTIONS: Logistic regression and a bootstrap resampling procedure were employed for model development and internal validation.
MEASUREMENTS AND MAIN RESULTS: Two hundred sixty-six (10.2%) had blood cultures positive for Candida spp. Mortality was significantly higher in patients with candidemia than in patients with bacteremia (47.0% versus 28.4%; P<.001). Administration of total parenteral nutrition, prior antibiotic exposure, transfer from an outside hospital or admission from a nursing home, mechanical ventilation and presence of a central vein catheter were independent predictors of candidemia while the lung as a source for infection was protective. The prediction rule had an area under the receiver operating characteristic curve of 0.798 (95% CI 0.77-0.82). Internal validation using bootstrapping technique with 1000 repetitions produced a similar area under the receiver operating characteristic curve of 0.797 (bias, -0.037; root mean square error 0.039). Our prediction rule outperformed previous rules with a better calibration slope of 0.96 and Brier score of 0.08.
CONCLUSIONS: We developed and internally validated a prediction rule for candidemia in hospitalized patients with severe sepsis and septic shock that outperformed previous prediction rules. Our study suggests that locally derived prediction models may be superior by accounting for local case mix and risk factor distribution.
PMID: 25813550 [PubMed - indexed for MEDLINE]