Algorithmic antibiotic decision-making in urinary tract infection using prescriber-informed prediction of treatment utility

Abstract Predicting antibiotic treatment outcomes could help tackle antibiotic resistance by guiding prescribing decisions. Existing approaches do not quantitatively incorporate the judgment of clinician users. Our antibiotic decision-making algorithm predicted treatment outcomes for 13 antibiotics using clinical prediction models trained on prescribing and urine culture data from 93,906 patients, then weighted outcomes using treatment decisions made by 49 clinicians in an antibiotic choice ranking exercise. In a simulation using Emergency Department data, the algorithm chose more correctly-targeted World Health Organization Access category antibiotics (75.6% of cases versus 11.9%, 95% confidence interval of difference 57.6% to 69.7%, p < 0.001) and oral antibiotics (69% versus 22.6%, 95% confidence interval of difference 39.5% to 53.4%, p < 0.001) than human prescribers, and fewer intravenous antibiotics (31.2% versus 65.8%, 95% confidence interval of difference −41.9% to −27.1%, p < 0.001). These results show that our algorithm could improve antibiotic prescribing decisions by combining human judgment with data-driven probability predictions.