Identifying patients at high risk for carbapenem-resistant Enterobacterales (CRE) carriage on admission to acute care hospitals: validating and expanding on a public health model

TL;DR

This study validates and improves a model to identify patients at high risk of carrying CRE bacteria upon hospital admission using healthcare data.

Contribution

The study expands a public health model for CRE risk prediction by incorporating healthcare system variables, improving its performance.

Findings

The original Chicago model had an AUC of 0.85 when validated using Atlanta public health data.

Adding healthcare system variables increased the AUC from 0.68 to 0.73 in the Atlanta clinical dataset.

The model performed well in a geographically distinct area and academic healthcare network.

Abstract

Validate a public health model identifying patients at high risk for carbapenem-resistant Enterobacterales (CRE) on admission and evaluate performance across a healthcare network.

Retrospective case-control studies

Adults hospitalized with a clinical CRE culture within 3 days of admission (cases) and those hospitalized without a CRE culture (controls).

Using public health data from Atlanta, GA (1/1/2016–9/1/2019), we validated a CRE prediction model created in Chicago. We then closely replicated this model using clinical data from a healthcare network in Atlanta (1/1/2015–12/31/2021) (“Public Health Model”) and optimized performance by adding variables from the healthcare system (“Healthcare System Model”). We frequency-matched cases and controls based on year and facility. We evaluated model performance in validation datasets using area under the curve (AUC).

Using public health data, we matched 181 cases to 764,408 controls, and the Chicago model performed well (AUC 0.85). Using clinical data, we matched 91 cases to 384,013 controls. The Public Health Model included age, prior infection diagnosis, number of and mean length of stays in acute care hospitalizations (ACH) in the prior year. The final Healthcare System Model added Elixhauser score, antibiotic days of therapy in prior year, diabetes, admission to the intensive care unit in prior year and removed prior number of ACH. The AUC increased from 0.68 to 0.73.

A CRE risk prediction model using prior healthcare exposures performed well in a geographically distinct area and in an academic healthcare network. Adding variables from healthcare networks improved model performance.

Figures