Abstract
BACKGROUND: Consensus-based large language model (LLM) ensembles might provide an automated solution for extracting structured data from unstructured text in echocardiography reports. METHODS: This cross-sectional study utilised 600 intraoperative transoesophageal reports (100 for prompt engineering; 500 for testing) randomly sampled from 7106 adult patients undergoing cardiac surgery at two hospitals within the University of Pennsylvania Healthcare System. Three echocardiographic parameters (left ventricular ejection fraction, right ventricular systolic function, and tricuspid regurgitation) were extracted from both the presurgical and postsurgical sections of the reports. LLM ensembles were generated using five open-source LLMs and four voting strategies: (1) unanimous (five out of five in agreement); (2) supermajority (four or more of five in agreement); (3) majority (three or more of five in agreement); and (4) plurality (two or more of five in agreement). Returned LLM ensemble responses were compared with the reference standard dataset to calculate raw accuracy, consensus accuracy, error rate, and yield. RESULTS: Of the four LLM ensembles, the unanimous LLM ensemble achieved the highest consensus accuracies (99.4% presurgical; 97.9% postsurgical) and the lowest error rates (0.6% presurgical; 2.1% postsurgical) but had the lowest data extraction yields (81.7% presurgical; 80.5% postsurgical) and the lowest raw accuracies (81.2% presurgical; 78.9% postsurgical). In contrast, the plurality LLM ensemble achieved the highest raw accuracies (96.1% presurgical; 93.7% postsurgical) and the highest data extraction yields (99.4% presurgical; 98.9% postsurgical) but had the lowest consensus accuracies (96.7% presurgical; 94.7% postsurgical) and highest error rates (3.3% presurgical; 5.3% postsurgical). CONCLUSIONS: A consensus-based LLM ensemble successfully generated structured data from unstructured text contained in intraoperative transoesophageal reports.