Document is current

Article History

Received: 7 December 2025

Accepted: 14 May 2026

First Online: 12 June 2026

Declarations

:

: The use of the LLMs was funded by the UCSF Versa project and by the Butte Lab at UCSF.

: No authors, Dana Ludwig, Michelle Wang, James Buchanan and Trang Trinh, have conflicts of interest to declare.

: No ethics approval was required for this study as it includes only HIPAA de-identified patient data.

: Not applicable.

: Not applicable.

: Access to de-identified patient data for this project is available at physionet.org under a database project named “Clinical Notes and LLM Output for Post-marketing Adverse Drug Event extraction with Large Language Model”. Access to the project requires “Credentialed” level of access.

: Software for the LLM model “o1” that matches the software used in this study is available on github at github.com/danaludwig1/ADE-extraction-with-LLM-OpenAI-o1 . In addition, updated software adapted to run the same prompts under LLM model GPT-5.2 is available on Physionet.org, under software project named “Source Code for Post-marketing Adverse Drug Event extraction with Large Language Model”.

: Dana Ludwig, Michelle Wang and James Buchanan contributed to the study conception and design. Data preparation, LLM model prompting and scripting and (Reviewer 1) validation were performed by Dana Ludwig. Implementation of the gold standard ADE annotation dataset (Reviewer 2) was performed by Michelle Wang. Validation of 70 LLM-generated ADEs not found in the gold standard (Reviewer 3) was performed by Trang Trinh. All authors read and approved of the final version.