Role of patient descriptors in predicting antimicrobial resistance in urinary tract infections using a decision tree approach: A retrospective cohort study

BACKGROUND In general practice, many infections are treated empirically prior to or without microbiological confirmation. Prediction of antimicrobial susceptibility could optimise prescribing thus improving patient outcomes. Decision tree models are a novel idea to predict AMR at the time of clinical presentation. This study aims to apply a prediction model using a decision tree approach to predict the antimicrobial resistance (AMR) of pathogens causing urinary tract infections (UTI) for patients over 65 years based on pre-existing routine laboratory data. METHODS Data were extracted from the database of the microbiological laboratory of the University Hospitals Galway (UHG). All urine results from patients over 65 years, their microbiological analysis and susceptibility (AST) results from January 2011 to December 2015 were included. The primary endpoint was culture result and resistance to antimicrobials (nitrofurantoin, trimethoprim, ciprofloxacin, co-amoxiclav, and amoxicillin) commonly used to treat UTI. A non-parametric regression tree analysis i.e. a decision tree model was generated with the 75% of the dataset (training set) and validated with the remaining 25% (test set). The model performance was evaluated measuring Area Under the Curve Receiver Operating Characteristic (AUC_ROC) curve. RESULTS A total of 99,101 urine samples of patients over 65 years were submitted for culture over the five years and 27% had significant bacteriuria (≥104 cfu/ml) and AST. The most common identified causative organisms were E.coli, Klebsiella spp. and Proteus spp. E.coli was more often resistant to amoxicillin (66%) followed by Proteus spp. (41%). Klebsiella spp. and Proteus spp. were more often resistant to trimethoprim (78% and 54% respectively). E. coli resistance to nitrofurantoin is low (<10%). The decision tree model showed an AUC-ROC score of 0.68 for culture and in between 0.60 to 0.97 for antimicrobial resistance of the pathogens, with the inclusion of patient's descriptors only. Including the uropathogen in the model did not change model performance. CONCLUSIONS The decision tree models using patient descriptors available at the time of presentation showed fair to excellent performance in predicting culture and antimicrobial resistance. The presented models provide an alternative approach to decision making on antimicrobial prescribing for UTIs. Increasing more predictors in the model could improve the model performance. Prospective data collection, validation and feasibility testing of the model including data from other laboratories will progress the practical implementation of similar models.