AI Spots Hidden Pancreatic Cancer Signals 16 Months Before Doctors See It

Pancreatic cancer remains one of oncology's most formidable challenges, with more than 85% of cases diagnosed after metastasis and five-year survival rates below 15% for advanced disease. Now, researchers at Mayo Clinic have demonstrated that artificial intelligence can identify subtle disease signatures on routine CT scans more than a year before clinical diagnosis—potentially transforming the detection timeline for a cancer that typically reveals itself only after spreading beyond treatment.

The AI model, called REDMOD (Radiomics-based Early Detection Model), achieved 73% sensitivity in detecting pancreatic cancer on scans that radiologists had originally interpreted as normal, with a median lead time of 475 days—approximately 16 months—before patients received their diagnoses. Published in the peer-reviewed journal Gut, the retrospective validation study analyzed contrast-enhanced abdominal CT scans from 493 patients in an independent test set, correctly identifying 46 of 63 pre-diagnostic cases while maintaining 81.1% specificity.

Outperforming Human Radiologists on Invisible Disease

The performance gap between REDMOD and human readers proved particularly striking in the study's head-to-head comparison. Two board-certified abdominal radiologists reviewing the same pre-diagnostic scans achieved pooled sensitivity of just 38.9%—less than half the AI's detection rate. The disparity widened further when examining scans taken more than two years before diagnosis: REDMOD maintained 68% sensitivity at this extended timeframe, while radiologists detected only 23% of cases.

This performance difference reflects REDMOD's fundamental approach. Rather than searching for visible tumor masses, the AI analyzes texture and structural radiomic patterns in pancreatic tissue—quantitative features imperceptible to the human eye but apparently indicative of early malignant transformation. The model demonstrated 90-92% concordance across serial imaging in patients with multiple scans, suggesting it detects stable biological signatures rather than transient imaging artifacts.

Lead researchers Sovanlal Mukherjee and Ajit H. Goenka developed REDMOD using a training cohort of 969 patients from multi-institutional datasets, then validated the model's specificity in both a separate multi-institutional cohort (81.3%) and a public NIH dataset (87.5%).

The Path from Retrospective Promise to Clinical Reality

Despite the compelling retrospective performance, REDMOD faces the critical question confronting all early detection AI: whether technical sensitivity translates into improved patient outcomes. The 81.1% specificity means approximately one in five flagged patients would be false positives in this low-prevalence disease—a rate that could generate substantial downstream testing, patient anxiety, and healthcare costs without corresponding benefit if earlier detection doesn't enable more effective intervention.

Mayo Clinic researchers are addressing this uncertainty through AI-PACED, a prospective clinical study that will deploy REDMOD in high-risk populations, including patients with new-onset diabetes and elevated pancreatic cancer risk scores. This trial will determine whether AI-flagged findings lead to actionable earlier diagnoses and improved survival, or primarily generate additional workup without outcome benefits.

REDMOD is not FDA-approved for clinical use or population screening. The technology analyzes standard-of-care imaging performed for other indications, potentially creating opportunities to extract additional diagnostic value from existing healthcare encounters without requiring specialized imaging protocols.

For a disease where late diagnosis remains the primary driver of mortality, REDMOD represents both technical validation that routine imaging contains early cancer signals and a test case for whether radiomics-based detection can bridge the gap between retrospective performance metrics and measurable clinical utility.

Sources

1. A Mayo Clinic AI re-analysed CT scans that had originally been read as normal — and detected the signature of pancreatic cancer in 73% of patients who later developed it, a median of 16 months before their diagnosis. — scienceblog.com