Artificial intelligence clinical decision support is increasingly used to support more consistent care, yet antimicrobial prescribing remains a difficult setting for automation. Switching from intravenous to oral antibiotics can be time-sensitive and is often made under uncertainty, with heterogeneous practice and differing risk perceptions. A mixed-method evaluation with UK prescribers assessed an AI-enabled clinical decision support system (CDSS) for intravenous-to-oral switching. Interviews and questionnaires were paired with an interactive clinical vignette experiment that compared standard-of-care decision making with AI and guideline-based recommendations. The focus was on how clinicians responded to advice, how often explanations were consulted and what factors shaped confidence in routine use.
Testing AI Support in Intravenous-to-Oral Switching
Forty-two clinicians from 23 hospitals in the UK participated between April 23, 2024 and Aug 16, 2024. Participants included consultants and training-grade doctors across infectious diseases, microbiology, pharmacy and other specialties, with varied familiarity with AI. In the vignette experiment, a custom web application presented 12 cases of patients receiving intravenous antibiotics. For each case, clinicians made a binary decision on suitability for switching to oral treatment.
Standard-of-care (SOC) materials included patient data displays and a UK Health Security Agency (UKHSA) intravenous-to-oral switch decision aid. In the decision-support condition, clinicians also received recommendations from an AI model and a guideline-based CDSS. Explanations were available via optional pages intended to add detail without overwhelming the main interface.
Must Read: Human–AI Collaboration Lifts Diagnostic Decisions
Across most vignettes, switching decisions did not differ meaningfully between SOC and decision support. Completion times were similar, with an overall mean of 148.5 seconds per vignette and no significant difference between conditions. Several vignettes still produced diverse decisions, underscoring that switching judgements can remain contested even when clinicians review the same information. Where differences emerged, decision support generally shifted clinicians towards not switching.
The clearest change appeared in one vignette where both the AI and the guideline CDSS recommended not switching and this matched the recorded action in the underlying dataset. In that scenario, most SOC participants chose to switch, whereas most participants exposed to decision support chose not to switch, with a significant difference in choices. When all 12 vignettes were considered together, the overall analysis across 504 observations indicated reduced odds of switching under decision support compared with SOC.
Clinician Behaviour and Explanations Used Infrequently
Interviews described antimicrobial prescribing as a blend of patient-specific judgement and guideline-led practice, shaped by behavioural and social influences. Participants highlighted cautiousness, resistance to change, personal biases and perceptions of risk linked to longer-term resistance. Technology use to support prescribing was uneven. Some clinicians reported using digital guidelines, while a substantial minority reported using no technology-based clinical decision support.
The vignette experiment suggested that clinicians did not accept AI advice uncritically. For vignettes designed with incorrect AI recommendations relative to the defined ground truth, participants often maintained their original decision rather than following the recommendation. This pattern indicated that in the controlled vignette setting AI outputs did not automatically produce over-reliance when presented alongside familiar SOC resources.
Explanations were available but rarely used at the point of decision. Just over half of participants accessed an AI explanation at least once, but across all vignettes where decision support was available, AI explanations were opened only 9% of the time. Explanations linked to the guideline-based CDSS were accessed even less frequently. When explanations were opened, they accounted for roughly a quarter of the time spent on the case, implying a trade-off between interpretability and speed.
Perceived influence did not fully align with observed behaviour. Half of participants disagreed that the AI CDSS changed their decision making, despite the overall finding that decision support reduced switching odds. This mismatch suggested that subtle shifts can occur without clinicians recognising a change in their own behaviour, particularly when decision support reinforces a cautious stance rather than prompting an overt reversal.
Usability, Trust and Adoption Conditions
Usability and acceptance measures were broadly positive. The system usability scale (SUS) score was 72.32 out of 100. Technology acceptance model (TAM) ratings were moderate-to-positive, with mean scores of 3.6 out of 5 for perceived usefulness, 3.8 out of 5 for perceived ease of use and 4.1 out of 5 for self-efficacy. Most participants reported comfort with AI in health care. Many described reassurance when their clinical decision aligned with recommendations and a tendency to investigate further when recommendations did not align.
Trust was described as conditional and evidence-led. Clinical evidence was repeatedly identified as central to confidence, alongside usability and clarity on what information the system uses to generate recommendations. Trust was also framed as something built over time through experience, with suggestions such as operating in shadow mode and providing training to support safe introduction. Reported barriers included behavioural inertia, limited readiness of hospital technological infrastructure and the uncertain nature of antimicrobial prescribing.
The broader context of antimicrobial resistance (AMR) shaped perceived relevance. AMR was described as a growing global threat, estimated to cause around 1.14 million deaths annually, with financial implications including costs of more than £180 million each year to the UK’s National Health Service. Within this context, two practical roles were emphasised: supporting clinicians who are not antimicrobial stewardship experts and helping pharmacists and microbiologists prioritise attention towards more complex patients.
An AI-enabled CDSS for intravenous-to-oral antibiotic switching was viewed as usable and potentially integrable by UK clinicians, but its influence on decisions was selective. Most vignette decisions were similar with or without decision support, clinicians often disregarded incorrect recommendations and the clearest shift occurred when both decision-support approaches encouraged not switching. Explanations were available but used infrequently, indicating that workflow fit and evidential confidence are likely to shape uptake. Prospective clinical trials were identified as necessary to assess clinical impact in real-world care.
Source: The Lancet Digital Health
Image Credit: iStock
References:
Bolton WJ, Wilson R, Gilchrist M et al. (2025) The impact of artificial intelligence-driven decision support on uncertain antimicrobial prescribing: a randomised, multimethod study. The Lancet Digital Health, 7(11):100912.
