Pancreatic cancer: an AI flags people at high risk from their medical records

Why is pancreatic cancer so hard to catch early?

The pancreas sits behind the stomach and helps with digestion and blood sugar. Its cancer gives few signs at first. According to the authors, around 80% of people are diagnosed once the disease has already spread, and very few then live another five years.

When it is found early, it can be cured with surgery, chemotherapy and radiotherapy. But this cancer is fairly rare, so screening everyone above a certain age would mean a lot of unnecessary tests. Doctors need a way to know whom to watch first.

What did the researchers do?

The team used two huge databases. In Denmark, the national patient registry, with 6 million people including 24,000 pancreatic cancer cases. In the United States, the records of military veterans, with 3 million people including 3,900 cases.

Their AI is a deep learning model, meaning a program that learns patterns by itself from many examples. It looks at no scans and no blood results: only the sequence of diagnoses recorded over the years (such as diabetes or digestive problems). It gives a risk score for the next 3 to 36 months.

To check that it was not just picking up last-minute signs, the researchers also trained it without the diagnoses recorded in the 3 months before cancer was found.

What were the results?

In Denmark, for cancer within 36 months, the best model reached an area under the curve of 0.88. This score runs from 0.5 (chance) to 1 (perfect). It fell to 0.83 when the last 3 months of data were removed, which shows the AI also spots older signals.

Among the 1,000 people over 50 judged most at risk, the estimated relative risk was 59: these people had around 59 times the risk of the general population.

The Danish model, used unchanged on US records, dropped to 0.71. It had to be retrained on US data to climb back to 0.78.

What could this change for patients?

The aim is not to diagnose cancer but to pick a small group of people who could be offered closer monitoring, such as regular scans. Today, this kind of monitoring mostly exists for people with a family history or a known genetic change.

If the approach holds up, it could widen that net to people who would never have been noticed, with no extra test at the start since their records already exist.

What are the limits?

This is a retrospective study: the AI was tested on past records, not by following patients in real life. We do not yet know whether monitoring the people it flags actually helps them live longer.

The big drop between Denmark and the United States shows that a model has to be adapted to each health system, because records are not filled in the same way. And a high score does not mean you have cancer: most people flagged will never develop it.