An AI model predicted pancreatic cancer risk up to three years before diagnosis using routine health records from nearly 40,000 patients.
The model analysed health histories collected over time alongside routine laboratory test results spanning an average of a decade or more.
Researchers found it could distinguish between people at higher and lower risk of pancreatic cancer three years before diagnosis, although further validation is under way.
Study co-author Cornelius Thiels, a surgical oncologist at Mayo Clinic in Rochester, Minnesota, said universal screening for pancreatic cancer was not feasible, prompting the researchers to explore whether routinely collected data could identify higher-risk patients.
“We know that pancreatic cancer forms over five to seven years, but the things that a clinician or patient sees don’t happen until it’s too late,” he said.
The Mayo Clinic research used data from 6,066 people with pancreatic cancer and 33,396 controls, whose clinical histories covered between 7.5 and 19 years.
Researchers developed the model to identify subtle patterns in routinely collected health information that could indicate increased risk years before diagnosis.
Pancreatic cancer is relatively rare but is often diagnosed at an advanced stage. The American Cancer Society estimates around 67,000 new diagnoses and 52,000 deaths in the US in 2026.
Researchers assessed how well the model distinguished between higher-risk and lower-risk patients using the area under the receiver operating characteristic curve, or AUROC.
The model achieved an AUROC score of 0.853 when predicting pancreatic cancer three years before diagnosis. A score of 1.0 represents perfect discrimination, while 0.5 is equivalent to chance.
It also recorded an area under the precision-recall curve of 0.712, indicating its ability to identify patients at risk while limiting false positives.
The model recorded a calibration slope of 1.08, a measure of how closely predicted risks matched what subsequently happened.
Lead study author Chris Varghese, a surgical data scientist at Mayo Clinic in Rochester, said: “Our model showed that a greater than 50 per cent risk of pancreas cancer predicted by our model indicated an 88 per cent likelihood of being diagnosed with pancreatic cancer in one year.”
The model is now being deployed on a research basis as the team moves from analysing existing records towards prospective testing in a clinical setting.
Thiels said: “We’re proving that we can move this from a retrospective research tool into our clinical environment and run it prospectively for validation.”
The researchers are also working to validate the model prospectively within Mayo Clinic and at a non-Mayo system, while developing more advanced machine learning approaches.