Mp69-04 Machine Learning Models Employed at the Time of Urine Specimen Collection Predict Antibiotic Resistance on Final Culture

You have accessJournal of UrologyInfections/Inflammation/Cystic Disease of the Genitourinary Tract: Kidney & Bladder I (MP69)1 May 2024MP69-04 MACHINE LEARNING MODELS EMPLOYED AT THE TIME OF URINE SPECIMEN COLLECTION PREDICT ANTIBIOTIC RESISTANCE ON FINAL CULTURE Glenn Werneburg, Jacob Knorr, Sean McSweeney, Alex Milinovich, Lyla Mourany, Daniel Rhoads, and Sandip Vasavada Glenn WerneburgGlenn Werneburg , Jacob KnorrJacob Knorr , Sean McSweeneySean McSweeney , Alex MilinovichAlex Milinovich , Lyla MouranyLyla Mourany , Daniel RhoadsDaniel Rhoads , and Sandip VasavadaSandip Vasavada View All Author Informationhttps://doi.org/10.1097/01.JU.0001008892.86171.de.04AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Antibiotic resistance is increasing rapidly. Urine culture is the gold standard laboratory assay for urinary tract infection diagnosis, but requires up to 72 hours to result. During the time period from culture collection to final result, empiric antibiotics are frequently initiated when there is clinical suspicion for urinary tract infection (UTI). Up to 30% of empiric antibiotic prescriptions require modification upon culture finalization based on inadequate coverage. Using machine learning, we aimed to develop interpretable algorithms based on patient factors to accurately predict antibiotic resistance at the time of culture collection, three days before the final culture sensitivities resulted at our institution. We hypothesized that algorithms would accurately predict sensitivity to antibiotics. METHODS: We identified urine cultures with available sensitivities from the electronic medical record over a 7 year period. Patient factors known or suspected to be associated with antibiotic resistance were collected from the EMR and used for model training. A series of machine learning algorithms was trained to predict antibiotic resistance. Areas under the receiver operating characteristic curve (AUC) and other model performance metrics were calculated based on bootstrapping. RESULTS: 177,773 cultures from 111,938 patients were identified and used for algorithm training and validation. Top-performing models were ensemble algorithms, and predicted trimethoprim-sulfamethoxazole (TMP-SMX) resistance with AUC 0.73 (Figure 1), and nitrofurantoin resistance with AUC 0.71. The most pertinent variables within each model were identified using a drop-out method. CONCLUSIONS: Machine learning models accurately predicted microbial resistance to first line antibiotics for urinary tract infection, approximately 72 hours prior to cultures resulting. Additional validation in outside institution datasets is warranted and ongoing. Results of the present study have important implications for reduction selection pressure for antibiotic resistance through more accurately targeted therapy. Upon validation, clinical implementation of the algorithms could reduce time to resolution of symptoms and healthcare expenditures. Download PPT Source of Funding: N/A © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e1119 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Glenn Werneburg More articles by this author Jacob Knorr More articles by this author Sean McSweeney More articles by this author Alex Milinovich More articles by this author Lyla Mourany More articles by this author Daniel Rhoads More articles by this author Sandip Vasavada More articles by this author Expand All Advertisement PDF downloadLoading ...