A machine learning model to predict optimal antibiotic use in hospital medicine patients.

Objective: Unnecessary and suboptimal antibiotic use causes adverse outcomes at both the level of individuals and health systems. Prospective audit and feedback, a core aspect of antibiotic stewardship program (ASP) efforts, reduces such use, but is inefficient in the absence of a pre-screening process. To address this, we used a machine learning approach to stratify antibiotic orders based on the likelihood of benefiting from ASP review, and to identify the factors most influential in the model's predictions. Design: Machine learning model developed using expert-labeled point-prevalence data. Setting: Single-center adult academic hospital. Participants: Hospitalized patients 18 years or older on internal medicine services between May 2021 and August 2022. Methods: Infectious disease experts assessed antibiotic orders for necessity and optimal use to create labels which were used as ground truth to train a machine learning model that uses automatically queried data from the electronic health record including vital signs, laboratory values, microbiological data, and medication administration records. Results: Our model achieved an area under the receiver operating characteristic curve of 0.89 for antibiotic necessity and 0.80 for optimal use. Model predictions were driven largely by markers of clinical instability, inflammation, and infection. Simpler clinical indices of infection had no predictive ability. Conclusions: We describe a model that predicts antibiotic necessity and optimal use using routinely available data that can be automatically aggregated from medical records. This makes it a promising option for early identification and intervention on orders most likely to benefit from ASP review.