Utility of a clinical risk factor scoring model in predicting infection with extended-spectrum β-lactamase-producing enterobacteriaceae on hospital admission

Abstract

Objective: To validate the utility of a previously published scoring model (Italian) to identify patients infected with community-onset extended-spectrum β-lactamase-producing Enterobacteriaceae (ESBL-EKP) and develop a new model (Duke) based on local epidemiology.

Methods: This case-control study included patients 18 years of age or more admitted to Duke University Hospital between January 1, 2008, and December 31, 2010, with culture-confirmed infection due to an ESBL-EKP (cases). Uninfected controls were matched to cases (3:1). The Italian model was applied to our patient population for validation. The Duke model was developed through logistic-regression-based prediction scores calculated on variables independently associated with ESBL-EKP isolation. Sensitivities and specificities at various point cutoffs were determined, and determination of the area under the receiver operating characteristic curve (ROC AUC) was performed.

Results: A total of 123 cases and 375 controls were identified. Adjusted odds ratios and 95% confidence intervals for variables previously identified in the Italian model were as follows: hospitalization (3.20 [1.62-6.55]), transfer (4.31 [2.15-8.78]), urinary catheterization (5.92 [3.09-11.60]), β-lactam and/or fluoroquinolone therapy (3.76 [2.06-6.95]), age 70 years or more (1.55 [0.79-3.01]), and Charlson Comorbidity Score of 4 or above (1.06 [0.55-2.01]). Sensitivity and specificity were, respectively, more than or equal to 95% and less than or equal to 47% for scores 3 or below and were less than or equal to 50% and more than or equal to 96% for scores 8 or above. The ROC AUC was 0.88. Variables identified in the Duke model were as follows: hospitalization (2.63 [1.32-5.41]), transfer (5.30 [2.67-10.71]), urinary catheterization (6.89 [3.62-13.38]), β-lactam and/or fluoroquinolone therapy (3.47 [1.91-6.41]), and immunosuppression (2.34 [1.14-4.80]). Sensitivity and specificity were, respectively, more than or equal to 94% and less than or equal to 65% for scores 3 or below and were less than or equal to 58% and more than or equal to 95% for scores 8 or above. The ROC AUC was 0.89.

Conclusion: While the previously reported model was an excellent predictor of community-onset ESBL-EKP infection, models utilizing factors based on local epidemiology may be associated with improved performance.