The design pitted three approaches against one another. Physicians in the clinician-in-the-loop group reviewed and edited LLM-generated drafts of discharge summaries and referral letters. Those in the clinician-only group drafted the same documents from scratch using matched templates. Unedited LLM drafts formed a third, LLM-only comparator. In total, seventeen physicians submitted 48 discharge summaries and 48 discharge referrals in the clinician-in-the-loop group, 54 of each document type in the clinician-only group, and 48 of each in the LLM-only group, giving the researchers a substantial body of documents to compare on both speed and expert-rated quality.

On timing, the results were suggestive rather than decisive. The clinician-in-the-loop workflow showed a shorter creation time for discharge summaries, with a beta of −59.4 seconds (95% CI, −150.6 to 31.7; P = .18), and a longer creation time for referrals, with a beta of 52.9 seconds (95% CI, −7.0 to 112.9; P = .08), compared with clinician-only drafting. Neither difference reached statistical significance, so the primary analysis showed no reliable time advantage for the AI-assisted approach. The direction of the estimates, however, hinted at potential time savings for certain document types, a pattern the authors note warrants further evaluation.

Quality told a clearer story. In most expert-rated quality domains, documents produced through the clinician-in-the-loop workflow outperformed those written by clinicians alone. The LLM-only drafts, meanwhile, were the fastest to produce but carried the lowest quality of the three groups. This pattern underscores a central concern that has shadowed the use of large language models in clinical documentation: hallucinations and factual errors remain a real risk, and raw AI output is not fit for clinical use without human scrutiny.

The trial’s conclusion is measured. A clinician-in-the-loop approach generally improved document quality, and although documentation time did not differ significantly in the primary analysis, the findings suggest the workflow may save time for some document types. The authors emphasize that active clinician supervision is essential and call for further evaluation of clinician-in-the-loop workflows in real-world clinical settings, beyond the simulated records and single ophthalmology department tested here. The trial was registered at ClinicalTrials.gov under identifier NCT07187050 on September 18, 2025.

For a field racing to deploy generative AI in hospitals, the pilot offers a sobering data point: the promise of AI-drafted clinical documents may lie not in replacing clinicians but in pairing fast machine drafts with careful human revision, a combination that in this trial produced higher-quality documents than either clinicians or the model achieved alone, even as the time savings remained unproven in this small, single-center setting.

Subject of Research: A pilot randomized controlled trial of an LLM-assisted clinician-in-the-loop workflow for clinical documentation time and quality

Article Title: Evaluating an LLM-assisted workflow for clinical documentation: A pilot randomized controlled trial on time and quality

Article References: Takayama, T., Sado, K., Suda, K., Tamura, H., Ueda-Arakawa, N., Ishihara, K., Ueda, Y., Okamoto, K., Santos, L. H. D. O., Oshika, T., Kuroda, T., Tsujikawa, A., & Miyake, M. (2026). Evaluating an LLM-assisted workflow for clinical documentation: A pilot randomized controlled trial on time and quality. PLOS Digital Health, 5(10), e0001799. https://doi.org/10.1371/journal.pdig.0001799

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001799

Keywords: large language models, clinical documentation, discharge summaries, referral letters, randomized controlled trial, clinician-in-the-loop, ophthalmology, Kyoto University Hospital, AI hallucinations, medical AI workflow, PROBE design, pilot trial