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

Large language models (LLMs) have been investigated for clinical documentation, with concerns about hallucinations and factual errors. Clinician review and revision of LLM-generated drafts are considered essential; however, the impact of such a workflow on both documentation time and quality remains unknown. This single-center, parallel-group, prospective, randomized, open-label, blinded-endpoint (PROBE) pilot trial evaluated whether a clinician-in-the-loop workflow improves clinical documentation time and quality compared with clinician-only drafting. From February 18 to March 14, 2025, at the Department of Ophthalmology, Kyoto University Hospital, 21 physicians were randomized, and 17 completed the intervention. Participants were randomized to either the Clinician-in-the-loop group or the Clinician-only group. Participants created discharge summaries and referrals for six simulated patient records. The Clinician-in-the-loop group reviewed and edited LLM-generated drafts, whereas the Clinician-only group drafted documents from scratch using matched templates. Unedited LLM drafts formed the LLM-only group. 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. Clinician-in-the-loop showed a shorter creation time for summaries (β = −59.4 seconds; 95% CI, −150.6 to 31.7; P = .18) and a longer creation time for referrals (β = 52.9 seconds; 95% CI, −7.0 to 112.9; P = .08) than Clinician-only, although neither difference was statistically significant. In most expert-rated quality domains, the clinician-in-the-loop workflow outperformed clinician-only drafting. LLM-only drafts were fastest but had the lowest quality. In conclusion, a clinician-in-the-loop approach generally improved document quality and, although the primary analysis showed no statistically significant difference in documentation time, suggested potential time savings for certain document types. These findings highlight the importance of active clinician supervision and support further evaluation of clinician-in-the-loop workflows in real-world clinical settings.Trial RegistrationClinicalTrials.gov Identifier NCT07187050 (registered on September 18, 2025).Author summary: Clinical documentation such as discharge summaries and discharge referrals is essential for patient care, yet it places a heavy burden on clinicians. Recently, large language models (LLMs) have been proposed as tools that might ease this burden, but it remains uncertain whether documents created with the help of an LLM can save time while still maintaining acceptable quality. In this study, we conducted a randomized controlled trial in which clinicians created clinical documents either by reviewing and editing drafts generated by an LLM or by writing them on their own. We also analyzed unedited LLM drafts to understand how the model performs without clinician review. We found that the clinician-in-the-loop approach generally improved document quality and may save time for some document types, although the primary analysis found no statistically significant difference in documentation time. In contrast, LLM-only drafts were produced most quickly but did not meet quality standards. These findings suggest that LLMs may support clinical documentation when paired with active clinician oversight. Such collaborative workflows warrant further evaluation in real-world clinical settings.