AI pancreatic cancer detection has produced a striking number: 475 days. That is the median lead time by which an artificial intelligence system called REDMOD identified pancreatic cancer on routine CT scans before patients were actually diagnosed, in a peer-reviewed study published in 2026. On the same scans, human radiologists spotted fewer than four in ten cancers. The result has been reported around the world, including in fresh coverage on 26 October 2026. But the most important fact about REDMOD is the one most headlines leave out: it is research, not a hospital test. No hospital can offer it today, no regulator has approved it, and the study itself says prospective validation must come first. This article explains what the research actually found, how accurate it really is, and why the gap between a study result and a clinic matters.
What REDMOD is
REDMOD stands for Radiomics-based Early Detection MODel. It was developed by researchers at Mayo Clinic in Rochester, Minnesota, with colleagues at the University of Texas MD Anderson Cancer Center and the University of Washington, and published in the journal Gut (BMJ) on 22 April 2026. The first author is Sovanlal Mukherjee; the senior author is Ajit Harishkumar Goenka of Mayo Clinic's radiology department.
Pancreatic ductal adenocarcinoma, the commonest form of pancreatic cancer, is notoriously hard to catch early. More than 85% of cases are diagnosed at an advanced, unresectable stage, and the US National Cancer Institute puts five-year survival below 15%. The disease is also projected to become the second-leading cause of cancer-related death in the United States by 2030. The reason is anatomical: early pancreatic tumours often cause no symptoms and no visible mass on a scan. REDMOD does not look for a visible tumour. It analyses standard abdominal CT scans — the kind already performed for other reasons — for subvisual radiomic signatures: subtle patterns in tissue texture, invisible to the human eye, that appear in the pancreas before a diagnosable cancer develops.
What the study actually did
The study was retrospective, meaning it looked backwards at scans already taken rather than testing patients in real time. Researchers collected contrast-enhanced abdominal CT scans obtained three to 36 months before a tissue diagnosis of pancreatic cancer. Crucially, every pre-diagnostic scan had been reported at the time as negative for any focal pancreatic mass, and a board-certified abdominal radiologist re-verified that no discernible mass was present. In other words, these were scans that had already been read as normal. Controls were patients whose scans showed no cancer and who had at least three years of follow-up without developing it, matched by age, sex, and scan date.
The core analysis used 1,462 CT scans: 219 pre-diagnostic cancer scans and 1,243 controls. A training set of 969 scans taught the model, and an independent test set of 493 scans — 63 cancer cases and 430 controls, simulating the low prevalence of the disease — measured how well it worked. Two further groups, 539 scans from multiple institutions and 80 scans from a public dataset, tested specificity only. Mayo's public summary of "nearly 2,000 scans" covers all of these cohorts together (2,081 in total); the core result rests on the 1,462.
Technically, REDMOD works in three steps. First, software automatically segments the pancreas on each scan. Then it extracts 968 quantitative radiomic features — mathematical descriptions of texture and pattern — and narrows them to a 40-feature signature, 90% of which are multi-scale wavelet-filtered textural features. Finally, an ensemble of machine-learning models (logistic regression, random forest, and XGBoost) combines those features into a risk score. The system assesses the whole pancreas as a diffuse "field effect" rather than pointing to a single spot, a limitation that matters later.
AI pancreatic cancer detection results: how accurate is REDMOD?
On the independent test set, REDMOD achieved an area under the curve (AUC) of 0.82, with sensitivity of 73.0% and specificity of 81.1%. In plain terms: it flagged about 73 in every 100 pre-diagnostic cancer scans, and correctly cleared about 81 in every 100 cancer-free scans. On the same scans, radiologists achieved sensitivity of just 38.9% — REDMOD roughly doubled the human detection rate, a difference the study reports as statistically significant.
The timing breakdown is where the result gets interesting. For scans taken more than 24 months before diagnosis, REDMOD's sensitivity was 68.0%, against 23.0% for radiologists — roughly three times higher. For scans taken three to 12 months and 12 to 24 months before diagnosis, sensitivity was 75.0% in both windows. The median lead time was 475 days, about 16 months. Of the 219 pre-diagnostic cases, 87 (40%) came from scans taken three to 12 months before diagnosis, 76 (35%) from 12 to 24 months, and 56 (25%) from more than 24 months — up to about three years. That maximum window is real, but it is not the typical result: the typical lead was 475 days, and headlines quoting "three years" without the median overstate what the study showed.
Two further results strengthen the finding. Test–retest stability was 90–92%, meaning the model gave consistent answers on repeat measurements. And specificity held up outside Mayo: 81.3% in the 539-scan multi-institutional group and 87.5% in the 80-scan public dataset. Wavelet-filtered features outperformed unfiltered ones (AUC 0.82 versus 0.74), supporting the idea that the signal lives in fine texture rather than in obvious structure.
The false-positive side must be stated alongside those numbers. A specificity of 81.1% implies that roughly 19 in every 100 cancer-free scans were flagged — in the test set, 81 of 430 controls were incorrectly flagged, as reported in secondary coverage consistent with that specificity. In a low-prevalence disease, that rate means many flagged people would not have cancer, and each flag would need follow-up imaging, anxiety, and possibly invasive procedures. That trade-off is central to why the system is not ready for general screening, and why several popular summaries that quote sensitivity without specificity are misleading.
Three different AI systems — do not mix them up
Much of the confusion around AI pancreatic cancer detection comes from merging three separate 2026 developments into one story. They use different data, report different metrics, and are at different stages.
REDMOD, described above, analyses CT scans and was published in Gut in April 2026. The second system is PANXEON, a blood test from City of Hope, published in Nature Medicine on 16 September 2026. It combines circulating microRNAs, exosomal microRNAs, and the CA19-9 marker into a single risk score. In 1,785 participants across the United States, Italy, Japan, and South Korea, it identified 86.8% of stage I–II pancreatic cancers, with false-positive rates of about 3.2% in low-risk controls and about 15.6% in high-risk controls, and it detected high-grade dysplasia in 64.3% of high-risk cyst cases. Its developers describe it as investigational, to complement rather than replace existing diagnostics.
The third is a health-records model from Mayo Clinic, presented at the American College of Surgeons Clinical Congress in Washington in September 2026. It uses longitudinal electronic health records and routine laboratory tests over a decade or more to estimate pancreatic cancer risk. In 6,066 people with pancreatic cancer, it reached a mean AUROC of 0.853 five years before diagnosis, with a positive predictive value of 88% at a risk threshold above 50%. This was a conference presentation, not a journal paper, and the reported control-group size differs between secondary sources, so this article does not publish a control figure for it. The three systems answer different questions — a scan signal, a blood signal, and a risk estimate — and their percentages cannot be compared directly.
Why hospitals cannot use REDMOD yet
The study's own conclusion is unambiguous: REDMOD is positioned for prospective validation in high-risk cohorts, described as a necessary step, and the authors state that prospective validation is paramount to confirm clinical utility. Mayo is advancing the system into a prospective study called AI-PACED (Artificial Intelligence for Pancreatic Cancer Early Detection), which will test AI-guided detection in patients at elevated risk and measure early detection, false positives, and clinical outcomes. A summary from BMJ itself states that REDMOD requires testing in high-risk patients — defined as those with unexpected weight loss and newly diagnosed diabetes — before it can be widely used in clinical practice.
No approval for REDMOD has been found in this research: no FDA clearance, no CE marking, no NHS adoption, and no routine hospital availability. If any article claims otherwise, it is going beyond the evidence. The practical problems are also unsolved. Because REDMOD assesses the whole pancreas rather than locating a lesion, a flagged scan shows no tumour to biopsy; follow-up would need further imaging or sampling to find disease that cannot yet be seen. Health systems are not currently set up to act on a pre-lesion AI flag — who gets flagged, how they are escalated when the scan looks normal, how risk is communicated, and who pays are all open questions.
That is also why general-population screening remains unrecommended. The US Preventive Services Task Force recommends against screening for pancreatic cancer in asymptomatic adults, a Grade D recommendation that predates REDMOD and is not changed by it. (That recommendation does not apply to certain high-risk inherited or familial groups, who are managed separately.) The study's limitations, stated by the authors themselves, reinforce the caution: it was not designed to evaluate performance across racial and ethnic groups, an important gap given known risk disparities; sensitivity has only been validated retrospectively, with public pre-diagnostic datasets for external sensitivity validation not yet available; and no survival benefit has been demonstrated — the study measured detection performance, not whether earlier flags save lives. Any future claim of lives saved must account for lead-time bias (earlier diagnosis appearing to extend survival even when it does not) and length-time bias (screening preferentially catching slower-growing disease).
What would have to be proven next
For REDMOD to move from research to clinic, several things must happen in order. First, prospective validation: running the system on new patients in real time, in high-risk groups, and measuring how it performs outside the historical dataset. Second, a demonstration that flagged patients can actually be followed up effectively — that further testing finds real disease without overwhelming services with false alarms. Third, evidence that earlier detection through this route improves outcomes that matter, not just detection statistics. And fourth, validation across diverse populations, so performance does not vary by who is being scanned. Until those steps are complete, the honest description of REDMOD is a promising, peer-reviewed research system — not a test your doctor can order.
That distinction should not diminish what was achieved. Finding a signal of pancreatic cancer in scans previously read as normal, a median of 475 days before diagnosis, at roughly double the sensitivity of expert radiologists, is a genuine scientific advance. It suggests the pancreas carries detectable texture changes long before a tumour is visible, which could one day reshape how high-risk patients are monitored. But the distance between a retrospective study and a screening programme is measured in prospective trials, not headlines — and the researchers themselves are the first to say so.
This article reports on published research. It is not medical advice. Anyone with symptoms, a family history of pancreatic cancer, or risk factors such as new-onset diabetes or unexplained weight loss should speak to a clinician. General-population screening for pancreatic cancer is not currently recommended.
Sources: Gut (BMJ), "Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability" (online first 22 Apr 2026, DOI 10.1136/gutjnl-2025-337266); Mayo Clinic News Network (29 Apr 2026); News-Medical / BMJ press summary (28 Apr 2026); The ASCO Post (May 2026); Becker's Hospital Review (19 Oct 2026); Nature Medicine PANXEON paper (16 Sep 2026) via Becker's Hospital Review and Knowridge; MedicalXpress and Drugs.com / HealthDay on the ACS Congress EHR model (Sep 2026); US Preventive Services Task Force pancreatic cancer screening recommendation. Full links: https://gut.bmj.com/content/early/2026/04/22/gutjnl-2025-337266 · https://newsnetwork.mayoclinic.org/discussion/mayo-clinic-ai-detects-pancreatic-cancer-up-to-3-years-before-diagnosis-in-landmark-validation-study/ · https://www.news-medical.net/news/20260428/AI-model-detects-pancreatic-cancer-years-before-clinical-diagnosis.aspx · https://ascopost.com/news/may-2026/ai-model-enables-earlier-detection-of-pancreatic-cancer-on-routine-ct-scans/ · https://www.beckershospitalreview.com/healthcare-information-technology/ai/ai-and-pancreatic-cancer-detection-4-notes/ · https://knowridge.com/2026/09/blood-test-detects-nearly-9-in-10-early-pancreatic-cancers/ · https://medicalxpress.com/news/2026-09-ai-pancreatic-cancer-years-diagnosis.html · https://www.drugs.com/news/acs-transformer-based-model-can-id-pancreatic-cancer-risk-131482.html
Related reading: Personalised mRNA Vaccine Cuts Melanoma Return Risk by 42% in World-First Trial · India Caps Prices of 137 Cancer Drugs — What Changes From November 9 · How Do Vaccines Work?