PREDICTION MODELS FOR CARBAPENEM-RESISTANT ENTEROBACTERIACEAE (CRE) AND OTHER MULTIDRUG-RESISTANT GRAM-NEGATIVE (MDRGN) BACTERIA IN HEALTHCARE SETTINGS
Goodman, Katherine Elizabeth
MetadataShow full item record
Background. Carbapenem-resistant Enterobacteriaceae (CRE) and other carbapenem-resistant organisms (CROs) pose urgent challenges to patient care. These bacteria are highly drug-resistant and are associated with significant attributable mortality. Current prevention strategies in United States (U.S.) healthcare facilities aim to reduce selective pressure from antibiotic exposure and to reduce patient-to-patient spread. These efforts are hampered by a lack of rapid and cost-effective diagnostics to identify these organisms. These diagnostic challenges leave basic epidemiological questions unanswered, including how many and which types of U.S. inpatients are asymptomatic carriers. Objectives. We aimed to measure the prevalence of, and risk factors for, CRO colonization among high-risk U.S. hospitalized patients and to develop statistical and machine learning prediction models that could help to address existing diagnostic limitations. Methods. To achieve these aims, we developed two study cohorts. The first, a one-year prospective cohort of Johns Hopkins Hospital (JHH) intensive care unit patients, screened patients for CRO carriage at unit admission. Isolates were speciated and molecularly characterized, and pre-admission exposure data were used to evaluate colonization risk factors and to develop predictive models of colonization with machine learning methodologies (Aim 1). The second, a retrospective cohort of JHH Gram-negative bacteremic patients, generated a clinical decision tree (Aim 2) and a risk score (Aim 3) to predict whether infections were extended-spectrum B-lactamase (ESBL)-producing. ESBLs confer resistance to most antibiotics except carbapenems, and rapid identification can reduce unnecessary carbapenem administration. Through the lens of this real-world example, we methodologically compared these two prediction approaches (Aim 3). Results. Aim 1 included 3,327 unit visits and 2,878 (87%) admission swabs. Our study found that 7.5% of patients were perirectally colonized with CROs and identified high organism and resistance mechanism diversity. Many variables were significantly associated with carriage, but resulting models were not highly predictive. Aims 2 and 3 analyzed 1,288 bacteremic patients and yielded higher performing prediction models for ESBL infection. We found that decision trees and risk scores performed similarly in our case study, but they offered different strengths and limitations. Conclusions. Statistical and machine learning prediction models offer an important complement to microbiological diagnostics. They can circumvent existing resource and practical constraints, but high biological heterogeneity can compromise their performance. Increasing familiarity with these methods, as well as refining distinctions between causal inference and prediction, may improve statistical tools for identifying colonization or infection with CROs and other multidrug-resistant bacteria.