Study design

In this study, we utilized a case–case–control design ^{Reference Kaye, Harris, Samore and Carmeli17} to identify risk factors for adult patients diagnosed with CRE infections admitted for >48 hours to a large, metropolitan, instructional medical center between January 2011 and December 2016. Cases of CRE infection were compared to patients diagnosed with carbapenem-susceptible Enterobacterales (CSE) infections and to controls randomly selected from among all patients admitted within 30 days of each CRE case. Patient records selected for this study were deidentified prior to analysis. ^{Reference Gallagher, Kuriakose, Haynes and Axelrod15,Reference Kaye, Harris, Samore and Carmeli17–Reference Rafailidis19}

Case definitions

Cases of CRE were defined as initial admission of adults admitted between January 2011 and December 2016 who were diagnosed with a bacterial infection of an Enterobacterales species resistant clinically or through sensitivity testing to carbapenems ≥48 hours after admission.

Cases of CRE were identified through a database maintained by the medical center’s antimicrobial stewardship program to track patients having an antibiotic-resistant infection. Generally, CRE infections are identified phenotypically based on carbapenemase production, ^{20–Reference Pierce, Simner and Lonsway23} or confirmed to harbor a known carbapenemase gene. ^{Reference Van Der Zwaluw, De Haan, Pluister, Bootsma, De Neeling and Schouls24}

Cases of CSE were selected from the same patient population as the CRE cases within a 30-day window for admission, with diagnostic pathogens identified as susceptible to carbapenems. Eligible CSE cases were independently identified by the medical center data warehouse and a subset was randomly selected for inclusion in the study.

Controls were defined as adult patients admitted to any service within a 30-day window from a CRE case for >48 hours who did not meet either of the above case definitions during that admission. All matched controls were randomly selected from the population of eligible patients by the information warehouse.

This project was determined to be exempt by The Ohio State University Institutional Review Board (study no. 2016E0677) and was granted a full waiver of HIPAA authorization by the HIPPA Privacy Board.

Data collection, management, and analysis

The following information was collected from the medical records of patients from all 3 patient groups: date of admission, age at admission, onset date of infection, sex, length of hospital stay (LOS), race, ethnicity, ZIP code, self-identified history of tobacco and alcohol use, Charlson score at admission, use of dialysis and catheter during the admission, prescription for antibiotics in the 90 days prior to admission, and organism with which the patient was infected.

The target sample size for this study was 110 per study group using α = 0.05 and β = 0.20 based on an expected prevalence of exposure to β-lactam antibiotics of 15% in controls.

Documented use of antibiotics for the previous 90 days prior to the date of admission was recoded as yes or no and the general class of antibiotics (β-lactams, others). The total number of antibiotics prescribed in the previous 90 days for any individual was recorded both by antibiotic class and total number.

The self-reported use of either alcohol or tobacco products on or nearest the date of admission was used to classify patients as alcohol drinkers (yes or no) or smokers (yes or no). Any positive response to use was recoded as a yes for analysis. The use of dialysis and catheters during admission or in the previous 90 days as well as any surgeries recorded in the medical record to have occurred within the previous 90 days were recorded.

The Charlson score was coded as a numeric score for analysis. ^{Reference Charlson, Pompei, Ales and MacKenzie25} The LOS was calculated as the number of days from admission to discharge. Age was recoded into 3 groups for analysis: 18–44 years, 45–64 years, and >65 years. ^{Reference Gallagher, Kuriakose, Haynes and Axelrod15} The ZIP code of the patient’s home address was recategorized into urban or rural using the 2013 Rural–Urban Continuum Codes from the US Department of Agriculture Economic Research Service. ²⁶

Data management and analysis were completed using Stata version 13.1 software (StatCorp, College Station, TX). Simple and multiple logistic regression models were based on previously identified risk factors related to CRE infection. ^{Reference Gallagher, Kuriakose, Haynes and Axelrod15} Univariable logistic regression models were analyzed to determine the unadjusted influence of independent variables on the outcome. Any clinically valid potential risk factors were reviewed independently and included in the full model if found to be plausible and statistically significant in the simple models. A full multivariable explanatory model was created using α = 0.10 as the initial inclusion criteria and α = 0.05 cutoff value for the final full model.

Backward stepwise multivariable logistic regression models were fit to develop the final model. For the multivariable model, an initial bivariable model was created with previous β-lactam use forced into the model as the primary factor of interest in this study and then adding all significant factors with an initial α < 0.10. A second model was then created using the significant variables from the initial model, retaining the previous use of β-lactams, and with entry and removal criteria lowered to 0.05 and 0.10, respectively. Finally, a model was developed that removed β-lactams specifically and added prior prescriptions for any class of antibiotic in the previous 90 days, using the same criteria for entry and removal of variables.