P-1594. Predictive models to improve antibiotic decision-making in the Emergency Department for Extended-Spectrum β-Lactamase-producing <i>Enterobacterales</i> infections

Abstract Background Extended-Spectrum β-Lactamase-producing Enterobacterales (ESBL-E) infections are associated with high morbidity and mortality, often due to delayed initiation of carbapenem therapy. Carbapenems must be used judiciously to prevent emergence of resistant organisms. We developed a machine learning algorithm (MLA) to identify patients at increased risk for ESBL-E infections in the emergency department (ED).Figure 1:Receiver Operating Characteristic (ROC) curve for predicting Extended-Spectrum β-Lactamase-producing Enterobacterales (ESBL-E) Receiver Operating Characteristic (ROC) curve for predicting Extended-Spectrum β-Lactamase-producing Enterobacterales (ESBL-E) infections in Emergency Department Patients using a gradient boosting machine learning model. The area under the curve (AUC) is 0.8, indicating good discrimination between patients with and without ESBL-E. Higher True Positive Rate (TPR) at a low False Positive Rate (FPR) signifies better model performance. Methods Adults (≥18) who visited an ED in the Johns Hopkins Health System between 1/2019 and 4/2024 and received Gram-negative therapies appropriate for sepsis or had documented ESBL-E infections were included. Timestamped ED vital signs, healthcare utilization, healthcare worker (HCW)-mediated patient connections, medication history, comorbidities and laboratory results populated prior to first antibiotic order were used as predictors; our primary outcome was ESBL-E infection defined by positive culture. The probability of ESBL-E infection was estimated using a gradient boosting machine learning algorithm (MLA). Because of the time-dependent nature of increasing ESBL-E infections, the MLA was derived using data from encounters prior to 7/1/2023 and evaluated in the remaining encounters using area under the operating characteristics curve (AUC) analysis.Table 1:Patient Cohort Demographics Results In total there were 129,773 patient encounters included, of which 4289 (3.3%) had a positive ESBL-E sample collected during their ED encounter. Out-of-sample predictive accuracy was high (AUC 0.80; Figure). Prior HCW-mediated connections to ESBL-positive patients and prior exposure (≤6 months) to carbapenems, cephalosporins and proton pump inhibitors were the strongest predictors of ESBL-E infection. Abnormal vital signs, increased age, and prior exposure to any antibiotic (prior 6 months) were also important factors in prediction accuracy. Conclusion An MLA using data readily available from the EHR at the time of antibiotic ordering provided good discrimination between ED patients with and without ESBL-E infection. Further refinement and implementation of MLA-driven decision-support in the EDs could drive more targeted use of broad-spectrum antibiotics and improved patient outcomes. Disclosures Jeremiah S. Hinson, MD, PhD, Beckman Coulter: Advisor/Consultant|Beckman Coulter: Grant/Research Support|Beckman Coulter: Stocks/Bonds (Private Company)