Due to its location deep in the abdominal cavity, its vague symptoms and a paucity of effective, practical and affordable screening technologies, pancreatic ductal adenocarcinoma (PDAC) is often diagnosed at an advanced stage and carries a very poor prognosis. There is therefore a critical need for practical and affordable approaches to screening. Ludwig Harvard researchers led by Asif Khan and Chris Sander described in a September issue of Cell Reports Medicine a potentially powerful approach for earlier detection of PDAC: a transformer-based AI model that learns interdependencies among events in patients’ longitudinal medical histories. Using longitudinal Veterans Affairs electronic health records from 19,426 PDAC cases and about 15.9 million controls, the model combined diagnostic and medication trajectories to predict PDAC risk within a 6-, 12- or 36-month window. In a cohort of 1 million patients, the top 1,000–5,000 highest-risk patients had a 3-year PDAC incidence 115- to 70-fold higher than a reference estimate based on age and sex alone. The AI-based model also identified diagnosis and medication exposures correlated with elevated risk, such as chronic inflammatory conditions and specific medications taken by patients. The model holds promise as a practical and relatively affordable way to improve early detection of PDAC.
Pancreatic cancer risk prediction using deep sequential modeling of longitudinal diagnostic and medication records
Cell Reports Medicine, 2025 September 16