2208. Development and Evaluation of Predictive Models for Estimating Infection Susceptibility to Empiric Treatment Regimens Among Patients with Pneumonia in Intensive Care Units

Abstract

BackgroundPredictive models for empiric antibiotic prescribing often estimate the probability of infection with multidrug-resistant organisms. In this work, we developed models to predict coverage of specific treatment regimens to better target antibiotics to high- and low-risk patients.MethodsWe established a retrospective cohort of adults admitted to the ICU in a 1,300-bed teaching hospital from November 1, 2011 to June 30, 2016. We included patients with a diagnosis of pneumonia and positive respiratory culture collected during their ICU stay. We collected demographics, comorbidities, and medical history from the electronic health record. We evaluated three penalized regression methods for predicting infection susceptibility to 11 treatment regimens: least absolute selection and shrinkage operator (LASSO), minimax concave penalty (MCP), and smoothly clipped absolute deviation (SCAD). We developed models for susceptibility prediction at two stages of the diagnostic process: for all pathogenic bacteria and for infections with Gram-negative organisms only. We selected final models based on higher area under the receiver operating characteristic (AUROC), acceptable goodness of fit, lower variability of the AUROCs in the cross-validation run, and fewer predictors.ResultsAmong 1,917 cases of pneumonia, 54 different pathogens were identified. The most frequently isolated organisms were: Pseudomonas aeruginosa (16.6%), methicillin-resistant Staphylococcus aureus (16.1%), and Staphylococcus aureus (13.5%). Frequently selected variables included age, Elixhauser score, tracheostomy status, recent antimicrobial use, and prior infection with a carbapenem-resistant organism. All final models used MCP or SCAD methods. Point estimates for the AUROCs in the training set ranged from 0.70 to 0.80, and estimates in the internal validation set ranged from 0.64 to 0.77.ConclusionMCP and SCAD outperformed LASSO. For some regimens, models predicted infection susceptibility with fair accuracy. These models have potential to help antibiotic stewardship efforts to better target appropriate antibiotic use.Disclosures All authors: No reported disclosures.