Mayo AI flags pancreatic cancer risk up to 3 years early

25 Sep 2026 (ET): American College of Surgeons released findings from Mayo Clinic researchers: an AI model reading longitudinal records and routine labs predicted pancreatic-cancer risk up to three years before diagnosis. Results go to ACS Clinical Congress 2026; not yet peer reviewed.

HEALTH desk — doctors do not screen everyone for pancreatic cancer, because it is uncommon, and most people who get it are found too late to cure. A score that uses only the medical chart and routine lab tests, and that claims years of warning, would matter if it holds, because those records already exist in ordinary hospitals. The American College of Surgeons says a program committee selected this abstract. It is not peer reviewed yet, so the three-year lead time is a meeting claim, not a checked paper.

What the release says the model is. The lede says researchers at Mayo Clinic have designed an artificial intelligence model that can potentially predict an individual’s risk of developing pancreatic cancer years before diagnosis. Potentially is the lede’s word. The line under the headline says Mayo Clinic researchers found the model reliably separated at-risk from low-risk individuals using routine health records. Reliably is that line’s word. Do not collapse potentially and reliably. The research will be presented at the American College of Surgeons Clinical Congress 2026 in Washington, Sept. 26–29, where the release says thousands of surgeons will convene. Will be presented is the release’s tense. The meeting had not started when this desk read the page on 25 Sep.

How deadly the release says this cancer is, and whose numbers those are. Pancreatic cancer is relatively rare but highly deadly, with about 67,000 new diagnoses and 52,000 deaths in 2026, according to the American Cancer Society. Those figures are the society’s, as the American College of Surgeons cites them. This desk did not recount the cancer registry. The release says the share of cancer deaths is outsized: pancreatic cancer accounts for about 3 percent of all new cancers but 8 percent of all cancer deaths. About, 3 percent, and 8 percent are the release’s.

Cornelius Thiels, as quotes, not as a clinic this desk walked. Thiels, DO, MBA, FACS, is a study co-author and a surgical oncologist at Mayo Clinic in Rochester, Minnesota. He said pancreatic cancer can be curable, but only when it is caught early, and fewer than one in five patients is diagnosed in time. He said survival for many patients is still measured in months, not years. Fewer than one in five, and months not years, are his sentences. He said universal screening for pancreatic cancer isn’t feasible, so the team set out to build an AI model that can identify patients at greatest risk. He said pancreatic cancer forms over five to seven years, but the things a clinician or a patient sees don’t happen until it’s too late. Five to seven years is his clock for how the cancer forms. It is not the model’s prediction window.

What the model read. The release says the model used each person’s longitudinal health history from the Mayo Clinic system. Longitudinal, here, means the chart over time, not a single visit. That gloss is this desk’s. The release calls that history the detailed, comprehensive patient information in the electronic health record. An electronic health record, often shortened to EHR, is the hospital’s digital chart. That gloss is this desk’s. The model combined that chart with results of routine laboratory tests obtained over an average of a decade or more. Routine laboratory tests are ordinary blood tests, not a special cancer scan. That gloss is this desk’s. A decade or more is the release’s average for those labs.

Who was in the dataset. The study included 6,066 individuals with pancreatic cancer and 33,396 controls with 7.5 to 19 years of clinical histories. A control, here, is a person in the comparison group who did not have the cancer. That gloss is this desk’s. Those two counts add to 39,462. That addition is this desk’s. The key takeaway on the page says the model drew on comprehensive health data from nearly 40,000 patients. Nearly 40,000 is the page’s round number for that total. The goal, Thiels said, was to identify subtle clues that could point to a risk of pancreatic cancer early on.

The three-year test, and what AUROC means. To test the model at predicting pancreatic cancer three years before diagnosis, the researchers calculated the area under the receiver operating characteristic curve, shortened to AUROC, to distinguish people at risk from people at low risk. The AUROC was 0.853. The release’s own scale: 1.0 would be perfect discrimination, and 0.5 would be no better than chance. Discrimination, here, means telling the two groups apart. That gloss is this desk’s. 0.853 is the researchers’ score on that three-year test. This desk did not rerun the model. The release also says the model showed a strong ability to identify patients truly at risk while limiting false positives, with an area under the precision-recall curve, shortened to AUPRC, of 0.712. A false positive is a healthy person the model flags anyway. That gloss is this desk’s. The release says the model was well calibrated, meaning the predicted risk lined up with what actually happened, with a calibration plot slope of 1.08. Well calibrated, and 1.08, are the release’s.

A one-year sentence that is not the three-year score. Chris Varghese, MBChB, lead study author and a surgical data scientist at Mayo Clinic in Rochester, said a greater than 50 percent risk of pancreas cancer predicted by the model indicated an 88 percent likelihood of being diagnosed with pancreatic cancer in one year. Greater than 50 percent, 88 percent, and one year are his sentence. The AUROC of 0.853 is the three-year test. Do not file the 88 percent as that three-year score.

What Varghese says would make it usable, still his tense. He said they built the model to be as generalizable, scalable, and easy to put into practice as possible. He said the data inputs are captured almost universally in hospital systems worldwide. He said if it is shown to work, it could be used in almost any setting. If, and could, are his words. Shown to work is still ahead of this abstract.

Where the team says the model is going. The researchers are deploying the model on a research basis, Thiels said. He said they are proving they can move it from a retrospective research tool into the clinical environment and run it prospectively for validation. Retrospective means the model looked back at charts of people already diagnosed. Prospective means watching people going forward, before anyone knows who will get the cancer. Those glosses are this desk’s. He said they are working to further validate the model within Mayo prospectively and, this year, at a non-Mayo system. This year is his phrase for that outside check. He also said they are developing more advanced machine learning architectures, which appear to improve the performance even more. Appear is his word. This desk did not see the prospective run or the newer architecture.

Who else is on the abstract, and what review it has had. Study co-authors with Thiels and Varghese are Leo Yan Li-Han, PhD; Tanios S. Bekaii-Saab, MD; Richa Bisht, MD; Ajit H. Goenka, MD; John D. Halamka, MD, MS; Ellen L. Larson, MD; Frank G. Lee, MD; Michael L. Kendrick, MD, FACS; Shounak Majumder, MD; Hojjat Salehinejad, PhD; and Mark J. Truty, MD, MS. The release says the authors have no disclosures to report. The citation on the page is Varghese C, et al., Enabling Digital Screening for Pancreatic Cancer using Artificial Intelligence Analysis of Disease Trajectories, Scientific Forum, American College of Surgeons Clinical Congress 2026. The ACS note on the same page says research abstracts presented at the Clinical Congress Scientific Forum are reviewed and selected by a program committee but are not yet peer reviewed. A program committee chooses what gets a slot at the meeting. Peer review is the later check by a journal. That gloss is this desk’s. Not yet peer reviewed is the ACS sentence. It is the limit on the scores above.

Who the about box says the college is. The American College of Surgeons is a scientific and educational organization of surgeons founded in 1913 to raise the standards of surgical practice. The about box says approximately 95,000 members, and that it is the largest organization of surgeons in the world. Approximately 95,000 is the college’s count. FACS designates a Fellow of the American College of Surgeons. This desk did not audit the membership roll.

What the card shows. The card is a labeled anatomical diagram of the human pancreas and the organs next to it, including the spleen, the duodenum of the small intestine, the bile ducts, and a close-up of a pancreatic islet with alpha cells, beta cells, and an exocrine acinus. It is an OpenStax textbook figure, Anatomy & Physiology, on Wikimedia Commons, file 1820 The Pancreas, under the Creative Commons Attribution 3.0 license. It is the organ this study is about. It is not a patient scan, not a Mayo product screen, and no date is printed on it. The catalog chip is HEALTH.

Plain English for the rest of the card. Mayo Clinic researchers told a surgeons’ meeting that a computer model, reading years of ordinary charts and routine blood tests, separated people who later got pancreatic cancer from people who did not, as far as three years out. They studied 6,066 people who had the cancer and 33,396 who did not. On that three-year test the ranking score, AUROC, was 0.853. One would be perfect. A coin flip is 0.5. The lead author also said that when the model put a person’s risk above 50 percent, that person had an 88 percent chance of a diagnosis within a year. That one-year line is a different claim from the three-year score. Pancreatic cancer is hard to catch early, and doctors do not screen everyone for it. The abstract was picked by a program committee. It is not peer reviewed yet. The team says it is now running the model as research, at Mayo and, this year, at a system that is not Mayo. The card is a textbook drawing of the pancreas.

PRIMARY here: the American College of Surgeons’ 25 Sep 2026 press release, Newswise stamp 25-Sep-2026 at 11:00 AM EDT, datelined Washington, with the same text on EurekAlert release 1145250 — Tier A PRIMARY, the college’s own release. STATUS PRIMARY. The desk label on the chip is CONFIRMED because that is the catalog word for a primary page we can stand on. The lede’s potentially, the subhead’s reliably, the Sept. 26–29 congress, the American Cancer Society figures of about 67,000 diagnoses and 52,000 deaths in 2026, the 3 percent and 8 percent shares, both Thiels quotes, the Mayo electronic-health-record history, the labs over an average of a decade or more, the 6,066 and 33,396 counts, the 7.5 to 19 years of histories, the three-year AUROC of 0.853, the AUPRC of 0.712, the calibration slope of 1.08, Varghese’s greater-than-50-percent and 88 percent in one year, the generalizable and if-shown-to-work sentences, the research-basis deployment, the prospective Mayo check and the non-Mayo check this year, the appear-to-improve line, the co-author list, no disclosures, the citation title, and the program-committee note that the abstract is not yet peer reviewed are that release’s. The 39,462 total is this desk’s addition of the two counts. The Sheila Evans contact is the EurekAlert page’s. NOT claimed: that this desk read the abstract, reran the model, or saw a chart; that the 88 percent is the three-year AUROC; that potentially and reliably are the same sentence; that the model is peer reviewed, FDA-cleared, or in routine screening; a result from the prospective run; a patient this desk identified; a stock tip, or investment advice. The card is the OpenStax pancreas diagram. Distinct from the already-filed fda-rasd-draft-guidance, nutshell-nts231-fda-ind, spring-health-vera-mh, and firefly-neurosigma-monarch. The catalog chip is HEALTH, not SIGNAL.