Most people with pancreatic cancer learn they have it when the disease is already spreading. Cornelius Thiels, a surgical oncologist at Mayo Clinic, says fewer than one in five patients are diagnosed in time for the cancer to be curable. On Sept. 25, his team announced a tool aimed at that gap: an AI model that reads a patient's own medical history and flags the risk up to three years before diagnosis.

Nearly 40,000 Records, One Question

The model was built by Thiels and Chris Varghese, a surgical data scientist at Mayo Clinic in Rochester, using electronic health records from the Mayo system. It combines each patient's long clinical history with routine laboratory results collected over a decade or more, according to the Mayo Clinic announcement.

The study set included 6,066 people who developed pancreatic cancer and 33,396 controls, each with between 7.5 and 19 years of records. The idea rests on biology. Thiels says pancreatic cancer develops over five to seven years, but neither doctor nor patient sees anything obvious until late. The model hunts for faint patterns in blood work and past diagnoses that hint at what is coming.

There is a useful parallel in a wearable monitor that found nighttime aldosterone surges which routine blood tests miss. Both stories make the same point: the data that would give an early warning often exists already, and nobody is reading it properly.

What the Numbers Actually Say

Mayo reports an AUROC of 0.853 when the model tries to separate future pancreatic cancer patients from controls three years before diagnosis. AUROC is a score where 0.5 is a coin flip and 1.0 is perfect discrimination, so 0.853 is good but not decisive.

The headline figure is stronger. Varghese said that when the model assigned a risk above 50%, 88% of those patients were diagnosed within a year. That sounds like a screening test. It is not one yet.

Why the 88% Figure Needs Care

In this study, about one person in seven had the cancer (6,066 of roughly 39,500). In an ordinary clinic population, pancreatic cancer is rare. A model tested on a dataset packed with cases will look sharper than it would when unleashed on everyone who walks into a GP's office. That is my reading of the design, not a criticism Mayo has made, and the team says prospective and external validation are still underway.

Thiels himself concedes that universal screening for pancreatic cancer is not feasible. The goal is narrower: pick out the small group of people worth a closer look. The model is currently deployed only on a research basis, Thiels said.

For readers, the takeaway is practical. Nobody should ask a doctor for this model, because it is not a clinical service. It is an argument that ordinary lab results carry more information than they currently deliver.

A Second Mayo Tool Reads the Scans

The records model is not Mayo's only bet. Earlier this year, the clinic published a separate imaging system, called REDMOD, in the journal Gut. It analyzes routine abdominal CT scans for tissue changes that appear before a tumor is visible. In that retrospective validation, Mayo said it identified 73% of cancers a median of about 16 months before diagnosis, nearly double the rate for expert radiologists reading the same scans without AI.

That work is moving into a clinical trial called AI-PACED. Ajit Goenka, a Mayo researcher, told NBC News that the trial has to follow participants for three to five years to see who develops cancer, so public availability is not close.

The two tools attack different halves of the problem. One decides who deserves attention. The other looks at what a scan shows. Whether they can be chained together, with the records model choosing who gets a CT and the imaging model reading it, is an obvious next question, though Mayo has not announced such a pairing.

Who Would Actually Get Screened

Any risk model has to answer an uncomfortable question: what happens to the people it flags? A false alarm means scans, biopsies and months of anxiety. A missed case means the current system continues unchanged.

The debate over heart calcium scans is a fair warning. That 10-year study of 6,000 adults found the extra test helped mainly patients at borderline risk, and added little for everyone else. An early-detection tool for pancreatic cancer will probably face the same reckoning: useful for a defined group, wasteful if offered to all.

What to Watch Next

Three results will decide whether this model matters. First, prospective validation, where the model predicts risk in patients whose outcomes are not yet known. Second, external testing at hospitals outside Mayo, whose patients and lab practices differ. Third, evidence that catching risk earlier leads to longer survival, not just an earlier diagnosis of the same outcome.

Until then, the 0.853 score is a promising start. Pancreatic cancer is one of the few cancers where even a modest head start could change what treatment is possible, which is why the research deserves attention and a long list of follow-up trials.