An artificial intelligence model developed at Mayo Clinic identified individuals at risk for pancreatic cancer up to three years before diagnosis using only routine health records and lab results. The findings, to be presented at the American College of Surgeons Clinical Congress 2026 in Washington, address a cancer that kills 52,000 Americans annually - largely because fewer than one in five patients receives a diagnosis when the disease is still curable.
The model drew on longitudinal health histories from 6,066 pancreatic cancer patients and 33,396 controls, spanning 7.5 to 19 years of clinical data. It achieved an AUROC score of 0.853 for distinguishing at-risk individuals from low-risk individuals three years before diagnosis, where 1.0 represents perfect discrimination. The area under the precision-recall curve reached 0.712, reflecting a strong ability to flag true cases while limiting false positives.
"Pancreatic cancer can be curable, but only when we catch it early - and fewer than one in five patients is diagnosed in time," said study co-author Cornelius Thiels, DO, MBA, FACS, a surgical oncologist at Mayo Clinic in Rochester, Minnesota. "As a result, survival for many patients is still measured in months, not years."
How the model reads subtle signals
Pancreatic cancer develops over five to seven years, but symptoms visible to clinicians and patients typically appear late. The Mayo team trained the model on the detailed, timestamped data already sitting in electronic health records - lab values, diagnoses, medication histories - looking for patterns too faint for a human reviewer to catch.
The calibration curve showed a slope of 1.08, meaning predicted risk tracked closely with actual outcomes. "Our model showed that a greater than 50% risk of pancreas cancer predicted by our model indicated an 88% likelihood of being diagnosed with pancreatic cancer in one year," said lead study author Chris Varghese, MBChB, a surgical data scientist at Mayo Clinic.
Built for broad deployment
The inputs the model requires are captured in hospital systems worldwide, which the researchers said was a deliberate design choice. "If it's shown to work, it could be used in almost any setting," Varghese said. The team is now moving the model from retrospective research into prospective validation within Mayo Clinic and, later this year, at an external health system.
Universal screening for pancreatic cancer is not currently feasible given the disease's low population-level incidence. A tool that narrows the screening pool to those with elevated risk could change the math. The researchers are also developing more advanced machine learning architectures that they said appear to improve performance further.
Why this matters for healthcare, science and research professionals
This study demonstrates a practical path from retrospective model to prospective clinical deployment using data that health systems already collect. For research scientists and clinical informaticists, the calibration rigor - AUROC, AUPRC, and calibration slope all reported - sets a benchmark for how prediction models in low-prevalence cancers should be validated. The external validation planned at a non-Mayo site this year will test whether the approach generalizes across different populations and EHR implementations, a step that often trips up academic models. Professionals working on clinical AI deployment can watch for those results to gauge real-world readiness.
Your membership also unlocks:
