Finding Pancreatic Cancer Clues in Routine Medical Records: Mayo AI Study Explores Risk Warnings Five Years Before Diagnosis

Early detection of pancreatic cancer presents a challenge: many patients already have advanced disease when diagnosed, yet screening everyone is difficult to implement. A Mayo Clinic research team is trying to identify people who warrant further evaluation using data collected during routine medical visits. According to the latest report from Medical News Today, this artificial intelligence model has been able to distinguish differences in risk using data from five years before diagnosis in a retrospective analysis, offering a potential pathway to earlier intervention that remains to be validated.

The study, led by Chris Varghese, was presented at the American College of Surgeons (ACS) 2026 Clinical Congress and has not yet been peer reviewed. The dataset released by ACS includes 6,066 patients with pancreatic cancer and 33,396 controls, each with 7.5 to 19 years of clinical records. The model combines electronic health records and routine laboratory tests to analyze trajectories of changes in health over time.

Study co-author Cornelius Thiels told Medical News Today that some components of the complete blood count were important signals for the model, while records of diabetes, pancreatitis, and other pancreatic diseases also contributed. These pieces of information alone are insufficient to predict cancer; the research seeks to identify how multiple signals intersect and change across years of medical records. If higher-risk groups can be reliably identified, subsequent testing could become more targeted.

However, the five-year results must be understood separately from the earlier version. In the analysis of data from three years before diagnosis released by ACS on September 24, the area under the receiver operating characteristic curve (AUROC) was 0.853. Medical News Today’s October 3 report cited an update from Varghese in which, using a slightly smaller sample, the AUROC for five years before diagnosis reached 0.853, while the updated three-year value was 0.76. AUROC measures the ability to distinguish people with different levels of risk and cannot be interpreted as “85.3% prediction accuracy.” Nor can the two sets of data be directly compared as points on the same timeline; full details of the methods and samples are still needed to clarify the differences.

The earlier ACS data also reported an area under the precision-recall curve (AUPRC) of 0.712 and a calibration slope of 1.08. Among people whose model-estimated risk exceeded 50%, 88% were diagnosed within one year. Medical Xpress republished this same set of ACS data; it does not represent a separate independent validation. In particular, the proportion of cases in the study sample is not equivalent to the incidence in the general population. These figures cannot be directly applied to routine outpatient care, and the model has not yet been shown to improve early detection or survival outcomes.

The team is already pursuing prospective validation at Mayo Clinic for research purposes and plans to evaluate the model in other health systems. This step will test whether the model can still reliably distinguish risk when applied to new patients, different testing practices, and different ways of recording medical histories, and whether its predicted probabilities correspond to actual disease occurrence. The existing conference results are insufficient to support routine clinical use.

Another key question is what to do after receiving a high-risk alert. Researchers are exploring how to connect these alerts with computed tomography, magnetic resonance imaging, endoscopic ultrasound, or blood biomarker testing, but the risk threshold for triggering testing has not yet been established. The real test is whether this pathway from a signal in medical records to further testing can detect cancer earlier while keeping false positives, the burden of testing, and patient anxiety reasonably controlled.