(AI)lephant in the Room: Utility of Generative AI to Develop Discussion Guides for Cardiac Surgery Morbidity and Mortality Conferences

OBJECTIVE: Morbidity & Mortality Conferences (M&Ms) are forums to assess adverse outcomes and to analyze their root causes. We sought to utilize a large language model (LLM) to generate discussion materials for cardiac surgery M&Ms, based upon data available in the electronic medical record (EMR). DESIGN: EMR data, including information from notes, bloodwork, vital sign monitoring, and the medication administration record, were input into a private instance of GPT-5. A prompt was iterated to accomplish three tasks, resulting in the generation of (1) a case outline, (2) a discussion table to identify deviations from evidence-based recommendations, and (3) discussion questions. Readability was assessed by Flesch-Kincaid Grade Level. SETTING: Our study was conducted within a high-volume cardiothoracic surgery department with residency- and fellowship-level training programs. PARTICIPANTS: A survey was administered to trainees in cardiothoracic surgery to assess perceptions of LLM-generated (LG) materials. RESULTS: EMR data was compiled for a total of fourteen cardiac surgery cases. LG and trainee-generated (TG) case outlines did not differ significantly in median word count (453.5 (LG) vs. 487.5 (TG); p = 0.68), though LG outlines had a higher median Flesch-Kincaid Grade Level (13.9 (LG) vs. 11.6 (TG); p = 0.003). 9 trainees completed a two-alternative, forced-choice survey comparing TG and LG outlines. LG outlines were more frequently preferred for presentation of baseline information (59.3% (16/27) (LG) vs. 40.7% (11/27) (TG)), while preferences on the basis of operative and perioperative information were more evenly distributed. LG discussion tables and questions had high perceived clinical applicability. CONCLUSIONS: An LLM effectively synthesized information from the EMR, in order to outline cases and prompt systems-based discussions about adverse events.