Predicting antimicrobial resistance at the emergency department for urinary tract infection

Abstract Background Urinary tract infection (UTI) is one of the most common infections observed among patients admitted to the emergency department (ER). Treatment failure due to antimicrobial resistance carries high risks of complications. Individual-level identification of pathogens and resistance is not available at the time of presentation, and so physicians initiate empiric treatment until culture data are available. This study aimed to identify factors related to the risk of harboring a resistant bacterial strain among patients admitted to the ER with suspected UTI, and to use these factors to develop a prediction model for resistance. Methods This retrospective case-control study was based on medical records from Assuta Ashdod Hospital, Israel (2017-2023). Multivariate models were fit to identify associations with resistance. Machine learning approaches were used for predicting resistance. Model performance was assessed by ROC analysis. Results In total, 6,823 patients were included, with mean age 68.9±21.4 years and two-thirds females. Resistance rates varied from 52% for ampicillin to 1% for amikacin. Key factors associated with resistance included age>80 (OR = 2.35, 95%CI: 1.77-3.16), prior resistance (OR = 5.17, 95%CI: 3.44-7.85), catheter use (OR = 4.29, 95%CI: 2.17-8.88) and recent antibiotic use (OR = 2.66, 95%CI: 2.19-3.23). The XGBoost algorithm performed best in predicting resistance, achieving AUC between 0.73 (95%Cl: 0.69-0.77) and 0.83 (95%Cl: 0.71-0.95). Conclusions Advanced age, male gender, recent antibiotic use, prior resistance, invasive sample collection, and long-term healthcare facility residence are major risk factors for antibiotic resistance in UTIs. A model based on individual patient information reliably predicts resistant UTI strains, potentially overcoming the limitations of traditional tools and enhancing empiric treatment decisions. Future research should focus on validating model effectiveness in real-world clinical settings. Key messages • Individual patient data can be used for personalized risk assessment for harboring a resistant bacterial strain, to help optimize utilization of antibiotic treatment. • Recent exposures to antibiotics and hospitalizations are key predictors for the acquisition of a resistant bacterial strain.