768. Use of a Machine-Learning-based Prediction Model to Guide Antibiotic De-escalation in the Treatment of Urinary Tract Infections

BACKGROUND: A patient-specific antibiogram (PS-ABG) issues personalized predicted antibiotic susceptibility results by incorporating patient factors into a prediction model. Predictions, reported as percent likelihood of susceptibility, are available to providers in real-time. In this study, we evaluated the performance characteristics of a PS-ABG based on a machine-learning algorithm in predicting susceptibility of Enterobacteriaceae isolated on urine cultures. METHODS: This cross-sectional study included 2,517 urine cultures with Enterobacteriaceae collected from 2,211 unique patients over a 12-week period from January 1 through April 15, 2019 in a single health system. Receiver operating curves (ROC) were generated for commonly prescribed antibiotics to assess discrimination. Threshold values to determine when an antibiotic could be used were then determined based on ROC curves. Brier scores were generated for all antibiotics collectively and for individual antibiotics to evaluate the accuracy of the predictions compared with that of the usual practice (UP) of traditional antibiograms. RESULTS: The ability of the PS-ABG to discriminate susceptible and nonsusceptible isolates varied by antibiotic [area under the curve (AUC) range: 0.71 - 0.95]. When all antibiotics were considered, AUC was 0.88 (95% C.I. 0.88 – 0.89). Brier score ranged from 0.0037 - 0.2087, representing between a 9 - 56% improvement compared with UP. For all antibiotics, the software had a 32% improvement over UP (median Brier score 0.0794 v. 0.1114, P < 0.0001). Overall, a susceptibility threshold of 95% was associated with a specificity of 96%. A threshold of 95% was associated with a ≥90% specificity in all agents except for cefepime (specificity 70%) and meropenem (specificity 73%). For cefepime and meropenem, specificity reached 90% at a threshold of 97%. CONCLUSION: The PS-ABG demonstrated excellent discriminatory power for all antibiotics tested and was more accurate than UP. A cutoff of 95% likelihood of susceptibility affords high specificity for most agents and may be a reasonable threshold for selecting an appropriate antibiotic. A higher susceptibility threshold yields similar specificity for cefepime and meropenem, but this finding is likely a result of the low number of resistant isolates. [Image: see text] DISCLOSURES: All authors: No reported disclosures.