Large Language Models Successfully Extract Data for Pancreatic Cyst Surveillance

A study evaluated the efficacy of large language models (LLMs) in extracting clinical variables from radiology reports, revealing high accuracy. An analysis of 3,198 scans from 991 patients demonstrated that LLMs could extract key elements related to pancreatic cysts with categorical variable accuracy reaching up to 99%. Continuous variable accuracy varied, with cyst size (92%) and main pancreatic duct size (97%). The findings suggest that LLMs can facilitate streamlined data curation for surveillance and future AI-based models.