Background. Rising antibiotic resistance increasingly compromises empiric treatment. Knowing the antibiotic susceptibility of a pathogen's close genetic relative(s) may improve empiric antibiotic selection.
Methods. Using genomic and phenotypic data from three separate clinically-derived databases of Escherichia coli isolates, we evaluated multiple genomic methods and statistical models for predicting antibiotic susceptibility, focusing on potentially rapidly available information such as lineage or genetic distance from archived isolates. We applied these methods to derive and validate prediction of antibiotic susceptibility to common antibiotics.
Results. We evaluated 968 separate episodes of suspected and confirmed infection with Escherichia coli from three geographically and temporally separated databases in Ontario, Canada, from 2010-2018. Across all approaches, model performance (AUC) ranges for predicting antibiotic susceptibility were greatest for ciprofloxacin (0.76-0.97), and lowest for trimethoprim-sulfamethoxazole (0.51-0.80). When a model predicted a susceptible isolate, the resulting (post-test) probabilities of susceptibility were sufficient to warrant empiric therapy for most antibiotics (mean 92%). An approach combining multiple models could permit the use of narrower-spectrum oral agents in 2 out of every 3 patients while maintaining high treatment adequacy (∼90%).
Conclusions. Methods based on genetic relatedness to archived samples in E. coli could be used to predict antibiotic resistance and improve antibiotic selection.