Bolton AI antimicrobial prescribing

The Impact of Artificial Intelligence-Driven Decision Support on Uncertain Antimicrobial Prescribing: A Randomised, Multimethod Study

Patient / Population Intervention / Exposure Comparison Outcome

In 42 UK clinicians from 23 hospitals who regularly prescribe antibiotics, randomised per case vignette, an AI-driven decision support system with explanations for intravenous-to-oral antibiotic switching did not differ from standard-of-care information alone for most switching decisions and completion times across 12 vignettes, but pushed decisions towards not switching where it did differ.

Design

Randomised multimethod vignette study, clinicians randomised, UK

Endpoint

Difference in clinicians' intravenous-to-oral switching decisions with vs without the AI CDSS, per vignette and overall; usability (SUS) and acceptance (TAM)

Relevance

ResultMost decisions and completion times equivalent; where a difference appeared the CDSS steered clinicians towards not switching (OR 0.13, 95% CI 0.03–0.50; P=0.0031). Explanations used only 9% of the time; SUS 72.3/100.

Bolton WJ, et al. Lancet Digit Health. 2025;7(11):100912. 10.1016/j.landig.2025.100912

Discussion & critique

A rare randomised look at how an AI CDSS actually shapes prescribing behaviour: clinicians could spot and ignore wrong advice, were swayed mainly towards continuing intravenous therapy, and almost never consulted the explanations — a sobering finding for explainable-AI design. Vignettes rather than patients, 42 self-selected clinicians recruited through personal networks, and no safety or outcome data.