Routine bloodwork and record patterns flagged pancreatic cancer risk years in advance
An artificial intelligence model built on longitudinal electronic health records identified people who later developed pancreatic cancer as early as five years before diagnosis. The work, presented at the American College of Surgeons Clinical Congress 2026 in Washington, D.C., used routine clinical data rather than specialized biomarker panels or dedicated screening scans, which is part of what makes the approach clinically interesting.
Pancreatic cancer remains difficult to catch early because its symptoms often arrive late and are nonspecific. By the time weight loss, jaundice, pain, or digestive complaints prompt urgent workup, the disease may already be advanced. A model that can separate higher-risk from lower-risk patients years earlier could shift attention toward a smaller group that might justify closer observation, targeted imaging, or additional laboratory evaluation.
The dataset was substantial: almost 40,000 patients from the Mayo Clinic health system, including 6,066 people who developed pancreatic cancer and 33,396 controls. Clinical histories spanned 7.5 to 19 years, giving the algorithm a long window to search for subtle changes across encounters, lab results, and diagnostic patterns. In practical terms, the model was not looking for a single red flag. It was learning from the rhythm of routine care.
Longitudinal clues mattered more than any single diagnosis
The model drew signal from complete blood count components, common pancreatic-related diagnoses such as diabetes and pancreatitis, and even patterns in otherwise unrelated healthcare interactions. That mix matters biologically. Pancreatic tumors can shape metabolism, inflammatory signaling, and hematologic markers long before a mass is obvious on imaging. Small shifts in these data streams may reflect an evolving disease process rather than a single dramatic event.
One especially striking example from the press release involved patients whose estimated risk exceeded 50 percent; 88 percent of them were diagnosed with pancreatic cancer within one year. In a setting where the disease is relatively uncommon, that type of enrichment could have major downstream implications for triage. It suggests the model may be able to concentrate suspicion in a subset of patients who already show a recognizable pattern in their records.
Performance metrics were also encouraging. Five years before diagnosis, the model achieved an area under the receiver operating characteristic curve of 0.853 in a slightly smaller cohort. Updated numbers were 0.84 at one year, 0.80 at two years, and 0.76 at three years before diagnosis. AUROC values range from 0.5, which is no better than chance, to 1.0, which represents perfect discrimination. The model also posted an area under the precision-recall curve of 0.712, a useful measure for uncommon outcomes such as pancreatic cancer. Calibration was reported with a slope of 1.08, suggesting predicted and observed risk were closely aligned in the research dataset.
This is not the same as a screening tool ready for clinic use.
Why calibration and thresholds will decide whether the model can help clinicians
A model can separate cases from non-cases reasonably well and still fail at the bedside if its predicted probabilities do not match real-world event rates. That is why calibration matters here. If a clinician is going to act on a risk estimate, the number has to mean something operationally, not just statistically. A patient predicted to have elevated risk should truly carry elevated risk at a level that justifies next steps.
The key question now is what threshold should trigger action. Should a flagged patient receive additional imaging, closer surveillance, a biomarker assay, or a referral to a specialist clinic? Each option carries tradeoffs. Computed tomography, magnetic resonance imaging, and endoscopic ultrasound can detect lesions more effectively than generic surveillance, but they also bring cost, resource burden, and the possibility of false positives. A model that casts the net too wide could create anxiety and unnecessary procedures; one that is too strict could miss the very cancers it aims to catch.
Investigators emphasized that the system was designed around information already present in everyday care, which is important because broad population screening for pancreatic cancer is not considered feasible. The disease is too rare for indiscriminate testing to make sense. A risk-stratified approach is more plausible, especially if the model can identify a small group in whom downstream tests become more efficient.
Prospective testing will determine whether the model changes care
The next phase moves beyond retrospective analysis. Prospective validation is underway within a healthcare system, and the team also plans to test the model in a non-Mayo setting. That step is essential because algorithms often perform differently outside the dataset that shaped them. Patient demographics, documentation habits, laboratory ordering patterns, and referral behavior can all alter performance once a tool leaves its development environment.
If the model holds up prospectively, the clinical workflow question becomes unavoidable: what happens after someone is flagged? Cornelius Thiels, DO, MBA, said the hope is that identifying a smaller, higher-risk population may make downstream screening tests more practical. He pointed to CT, MRI, endoscopic ultrasound, and blood-based biomarker tests as possible follow-up tools, but the right sequence still needs to be defined and tested prospectively.
Chris Varghese, MBChB, described the central value of the model as a chance to move detection upstream, from reacting to symptoms to identifying people who may benefit from earlier evaluation. That idea fits the biology of pancreatic cancer, which often evolves quietly over years. The challenge is making the signal specific enough to justify action before the disease becomes clinically obvious.
For now, the study offers a sophisticated retrospective proof of concept rather than evidence of improved survival, earlier stage at diagnosis, or better patient outcomes in routine care. Those endpoints have not been established. The model’s real-world value will depend on whether prospective testing shows that it can guide follow-up in a way that improves detection without flooding clinics with false alarms or unnecessary imaging.