Mayo Clinic's REDMOD AI Spots Pancreatic Cancer 16 Months Before Specialists, Landmark Study Finds

Mayo Clinic researchers have shown that an artificial-intelligence model trained on routine abdominal CT scans can identify pancreatic cancer up to three years before patients are formally diagnosed, in a multi-site validation study published in the BMJ journal Gut. The system, named REDMOD (Radiomics-based Early Detection Model), flagged 73% of prediagnostic cancers at a median lead time of roughly 16 months — nearly double the sensitivity of specialist radiologists asked to re-read the same images.

REDMOD looks for subtle texture changes in the pancreas — patterns invisible to the human eye that suggest cellular reorganization months or years before a tumor is large enough to be called cancer. In the validation cohort, the model analyzed nearly 2,000 CT scans, including 219 patients who were later diagnosed with pancreatic ductal adenocarcinoma but whose initial imaging had been read as normal by clinicians. The system maintained 87.5% specificity on external datasets, an unusually high bar for a screening tool aimed at a disease that affects only a small fraction of patients.

The performance gap with human readers widened as the lead time grew. For cases identified more than two years before diagnosis, REDMOD was accurate 68% of the time versus 23% for radiologists reviewing the same studies — roughly three times the hit rate. Pancreatic cancer kills more than 90% of patients within five years of diagnosis, in large part because most tumors are caught only after they have spread, so a tool that pushes detection back by a year or more could meaningfully shift the survival curve. Lead investigator Mayo Clinic radiologist Dr. Ajit Goenka framed the work bluntly: 'Is there information already sitting in routine CT scans that we are failing to extract?'

Mayo is now moving REDMOD into a prospective clinical trial, dubbed AI-PACED (Artificial Intelligence for Pancreatic Cancer Early Detection), that will study how the model performs when it is built into the workflows clinicians actually use — initially focused on high-risk groups such as patients with new-onset diabetes, a known early signal of pancreatic disease. The next test is regulatory: validating retrospective imaging data is one thing, but persuading the U.S. FDA and payors that a screening AI changes outcomes in live practice is what determines whether REDMOD becomes part of standard care or another promising-but-shelved algorithm.