Enhancing EHR-based pancreatic cancer prediction with LLM-derived embeddings

Pancreatic cancer (PC) is often diagnosed late, as early symptoms and effective screening tools are lacking, and genetic or familial factors explain only ~10% of cases. Leveraging longitudinal electronic health record (EHR) data may offer a promising avenue for early detection. We developed a predictive model using large language model (LLM)-derived embeddings of medical condition names to enhance learning from EHR data. Across two sites-Columbia University Medical Center and Cedars-Sinai Medical Center-LLM embeddings improved 6-12 month prediction AUROCs from 0.60 to 0.67 and 0.82 to 0.86, respectively. Excluding data from 0-3 months before diagnosis further improved AUROCs to 0.82 and 0.89. Our model achieved a higher positive predictive value (0.141) than using traditional risk factors (0.004), and identified many PC patients without these risk factors or known genetic variants. These findings suggest that the EHR-based model may serve as an independent approach for identifying high-risk individuals.