A tool that could identify high-risk patients years before symptoms appear
Among the cancers that most reliably elude human vigilance, pancreatic cancer has long stood apart — silent, swift, and almost always discovered too late. Researchers at Mayo Clinic have now trained an artificial intelligence system called Redmod to read what the human eye cannot: faint patterns in CT scans that precede a diagnosis by more than a year on average. In a domain where timing determines survival, this development suggests that the boundary between detection and fate may, at last, be movable.
- Pancreatic cancer kills nearly nine in ten patients within five years precisely because it produces no warning until it has already spread — making early detection not a convenience but a matter of life and death.
- Redmod identified 73% of future cancer cases in CT scans that radiologists had read as normal, and the gap widened further for scans taken more than two years before diagnosis, exposing the limits of unaided human perception.
- The AI's ability to correctly clear more than 80% of scans from people who never developed cancer is critical — without that specificity, the system would flood clinics with false alarms and erode the trust needed for adoption.
- Researchers are tempering enthusiasm with discipline: Redmod has only been tested retrospectively, and real-world prospective trials must confirm whether earlier flagging actually translates into lives saved.
- If validated, the tool could redirect clinical attention toward high-risk patients — those with unexplained weight loss or new-onset diabetes — creating an intervention window that currently does not exist.
Pancreatic cancer kills with a particular cruelty: it hides. Tumours grow silently, producing no symptoms until the disease has spread beyond the reach of surgery. This is why the five-year survival rate sits at roughly one in ten globally, and why more than 85 per cent of cases are diagnosed only after the disease has advanced past the point of meaningful treatment.
Researchers at Mayo Clinic and their collaborators may have found a way to change that arithmetic. Their artificial intelligence system, Redmod, can detect pancreatic cancer in routine CT scans an average of 475 days before a patient receives a formal diagnosis — identifying patterns too subtle for the human eye to register. The findings, published in the journal Gut, suggest a fundamental shift: catching the disease not as a crisis to be managed, but as a risk to be identified years in advance.
The numbers are striking. Trained and tested on scans from more than 1,400 people, Redmod correctly identified 73 per cent of cases that radiologists had read as normal. Those same radiologists caught only 39 per cent. For scans taken more than two years before diagnosis, the gap widened further — 68 per cent for the AI versus 23 per cent for doctors. Crucially, the system also correctly cleared more than 80 per cent of scans from people who never developed cancer, a measure of precision that guards against false alarms.
The stakes are clear. Modelling suggests that increasing the share of cases caught while still localised — from 10 per cent to 50 per cent — could more than double survival rates. Redmod still requires prospective real-world testing before clinical adoption, and whether it will improve outcomes when deployed on patients not yet diagnosed remains to be proven. But the potential is significant enough that the medical community is watching closely, sensing that one of medicine's most stubborn failures may finally be within reach of a different answer.
Pancreatic cancer kills with a particular cruelty: it hides. Tumours grow silently, producing no symptoms until the disease has already spread beyond the reach of surgery. By the time a patient feels sick enough to see a doctor, the cancer has usually won. This is why the five-year survival rate sits at roughly one in ten globally, and why more than 85 per cent of cases are diagnosed only after the disease has advanced beyond the point where meaningful treatment is possible.
A team of researchers at Mayo Clinic and their collaborators may have found a way to change that arithmetic. They have developed an artificial intelligence system called Redmod that can detect pancreatic cancer in routine CT scans an average of 475 days before a patient receives a diagnosis. The system does this by identifying patterns in the images that are too subtle for the human eye to register—changes so faint that radiologists reviewing the same scans miss them entirely. The findings, published on April 28 in the journal Gut, suggest a fundamental shift in how pancreatic cancer might be caught: not as a crisis to be managed, but as a risk to be identified years in advance.
The implications are stark. Redmod was trained and tested on scans from more than 1,400 people, including 219 patients whose earlier imaging had been read as normal but who later developed pancreatic cancer. In direct comparison, the AI system correctly identified 73 per cent of these cases. Radiologists reviewing the identical images caught only 39 per cent. The gap widened dramatically for scans taken more than two years before diagnosis: the system detected 68 per cent of cases, while doctors spotted 23 per cent. The model also proved robust across different hospitals and different scanner manufacturers, and it correctly classified more than 80 per cent of scans from people who never developed cancer—a crucial measure of specificity that prevents false alarms.
Why does this matter so much? Because timing is everything in pancreatic cancer. Modelling studies suggest that if the proportion of cases caught while still localised—before spread—could be increased from the current 10 per cent to 50 per cent, survival rates would more than double. A tool that could identify high-risk patients years before symptoms appear would create an entirely new window of opportunity. Patients with unexplained weight loss, new-onset diabetes, or other risk factors could be flagged for closer monitoring and earlier intervention, potentially catching tumours when surgery or other aggressive treatments still have a chance to work.
The researchers are careful about what they claim. Redmod needs prospective testing in real-world screening settings before it can be adopted into routine clinical practice. The system has shown what it can do in retrospective analysis—looking back at scans that were already taken. Whether it will actually improve patient outcomes when deployed prospectively, flagging people who have not yet been diagnosed, remains to be proven. But the potential is significant enough that the medical community is watching closely. If confirmed, this tool could transform pancreatic cancer from a disease that is almost always caught too late into one where early detection becomes possible, and survival becomes something more than a statistical rarity.
Notable Quotes
This temporal window holds profound significance, as attaining such early detection would substantially augment the probability of cure and improve survival— Mayo Clinic researchers
The timing of diagnosis is the single most critical determinant of survival outcomes— Researchers citing modelling studies
The Hearth Conversation Another angle on the story
Why is pancreatic cancer so hard to catch in the first place?
It doesn't announce itself. There are no early symptoms, and the tumours don't show up on imaging until they're already large and have often spread. By then, surgery isn't an option anymore.
And this AI system—Redmod—it's looking at the same CT scans a radiologist would see, but finding things the radiologist misses?
Exactly. It's detecting patterns so subtle that human eyes can't register them. The system was trained on thousands of scans, including ones from people who later developed cancer. It learned to see the whispers before the shout.
The numbers are striking. 73 per cent detection versus 39 per cent for doctors. But what does that actually mean for a patient?
It means a potential 475-day head start. Imagine being told about a cancer risk more than a year before you'd ever feel sick. That's time to pursue surgery, to catch it while it's still contained.
Does the system work the same way everywhere, or is it dependent on the hospital or the equipment?
It performed consistently across different hospitals and different scanner manufacturers. That's crucial—it means it's not just a quirk of one institution's setup.
What's the next step? Can doctors start using this tomorrow?
No. It needs real-world prospective testing first. They need to deploy it in actual screening settings and see if it actually improves survival rates, not just whether it can spot patterns in old scans.
And if it works in those trials?
Then you're looking at a fundamental change in how pancreatic cancer is managed. Instead of waiting for symptoms, you're identifying people at risk years in advance. That's the difference between a disease that kills nine out of ten people and one where survival becomes possible.