NCT07187050

LLM-Assisted vs Manual Writing for Clinical Documentation: Effects on Time and Quality

Evaluating an LLM-Assisted Workflow for Clinical Documentation: A Pilot Randomized Controlled Trial on Time and Quality

• Status: Completed
• Phase: N/A
• Study type: Interventional
• Enrollment: 21 (actual)
• Sponsor: Kyoto University, Graduate School of Medicine · Academic / Other
• Sex: All
• Age: —
• Healthy volunteers: Not accepted

Summary

The goal of this clinical trial is to learn whether an LLM-assisted writing workflow can reduce the time to complete hospital discharge summaries and discharge referrals and maintain or improve document quality compared with writing from scratch by clinicians. The study used six simulated patient records (no real patient data). The main questions it aims to answer are: * Does the LLM-assisted writing workflow reduce the time needed to complete each document compared with manual writing? * Does the LLM-assisted writing workflow improve (or at least maintain) document quality compared with manual writing, as rated by blinded experts? Researchers will compare LLM-assisted versus manual writing to see if the LLM-assisted approach is faster and has equal or better quality. LLM-only drafts (unedited first drafts) will be evaluated as a separate third group to understand the baseline quality of LLM output without clinician edits. Participants will create two documents-a discharge summary and a discharge referral-for each of six simulated cases. Those assigned to CocktailAI \& Modification group will use an LLM assistant (called CocktailAI) to generate a first draft for each document and then review and edit it to finalize; those assigned to the control group will write each document from scratch without LLM assistance.

Conditions

• Clinician-in-the-loop
• Clinical Documentation
• Large Language Model

Interventions

Type	Name	Description
OTHER	Template-Based LLM Assistant	This study uses CocktailAI, a template-based LLM assistant co-developed by the Department of Ophthalmology and Visual Sciences, Kyoto University Graduate School of Medicine, and Fitting Cloud Inc. (Kyoto, Japan). It is designed to extract relevant information from EHRs using LLMs and embed the extracted content into predefined templates. In this trial, the inputs are six simulated patient records (no real patient data). Text generation uses Gemini-2.0-flash-lite. Templates for discharge summaries and discharge referrals are pre-defined by a team member.
OTHER	Manual Writing	The same document templates are provided; however, all LLM instruction prompts are removed in advance. Clinicians manually write the documents, following the template structure, for each of the six simulated cases.

Timeline

• Start date: 2025-02-18
• Primary completion: 2025-03-14
• Completion: 2025-07-16
• First posted: 2025-09-22
• Last updated: 2025-09-22

Locations

1 site across 1 country: Japan