An artificial intelligence model predicts pancreatic cancer three years before diagnosis

The results of this research were presented at the 2026 American College of Surgeons Clinical Congress in Washington, September 26-29.

It is known that pancreatic cancer is relatively rare, but very deadly, as it is often diagnosed in advanced stages because its symptoms are not clear at first.

“Pancreatic cancer may be curable, but only when we catch it early, and in less than one in five patients it is diagnosed in time,” says Cornelius Theles, a surgical oncologist at the Mayo Clinic and one of the paper’s co-authors. “As a result, many patients’ survival after diagnosis is still measured in months rather than years.”

How does the model work?

Universal screening for pancreatic cancer isn't possible, Thiels says, so his team sought to develop an artificial intelligence model that identifies patients most at risk. He explains: "We know that pancreatic cancer forms over five to seven years, but what the doctor or patient sees does not happen until it is too late."

The model, developed by Thiels, lead author Chris Varghese, and their team, used patients' longitudinal health records from the Mayo Clinic system and combined them with the results of routine laboratory tests collected over an average of a decade or more.

The study's data set included 6,066 people with pancreatic cancer and 33,396 people in a control group with clinical records spanning 7.5 to 19 years, with the aim of identifying hidden signals that may indicate risk at an early stage.

To confirm the model's ability to predict pancreatic cancer up to 3 years before the actual diagnosis, the researchers used three different metrics, each answering a specific question about the model's performance.

The first method answers the question: Does the model differentiate between a patient at risk and someone else? It obtained a score of 0.853 out of 1.0, meaning that it is very close to perfection in differentiating between those who will get cancer and those who will not.

The second method answers the question: When the model warns of danger, is it right? It received a score of 0.712, meaning that its warnings are reliable most of the time, and it does not raise many false alarms.

The third method answers the question: When the model says that the risk surrounding a person is 50%, is this number correct? It received a score of 1.08 (the ideal number is 1.0), meaning its numbers are very close to reality.

“If the model says someone's risk is more than 50%, then there is an 88% chance they will actually be diagnosed with pancreatic cancer within a year,” says Chris Varghese, the study's lead author. That is, artificial intelligence can predict the disease three years before diagnosis. The higher the risk identified by the model, the more accurately it predicts infection over the following year. In other words, the prediction of infection within a year reflects how accurate the model is when the risk is high.

According to the researchers, the model is accurate in differentiating between those at risk and those who are not at risk, has few errors in its warnings, and its numbers are close to reality. This makes it a promising tool for early detection of pancreatic cancer.

“We built this to be as generalizable, scalable, and practical as possible,” Varghese asserts. Pointing out that the data inputs on which the model depends are collected almost comprehensively from hospital systems around the world. “If it proves to work, it can be used in almost any environment,” he adds.