The 475-Day Head Start: AI Spots Pancreatic Cancer Before Any Human Can

TL;DR

• Mayo Clinic researchers have published a validated AI model — REDMOD — that detected pancreatic cancer on routine CT scans with 73% sensitivity at a median of 475 days before clinical diagnosis.
• Pancreatic cancer kills ~92% of patients within five years because it's almost always found too late. This model finds it before symptoms appear, on scans taken for other reasons.
• The study, published in Gut (BMJ) on May 4, is retrospective — it hasn't been prospectively validated yet. But the signal is strong enough that prospective trials are now a clinical imperative, not a research curiosity.
• Separately, Caris Life Sciences launched an AI-powered test on May 5 that predicts both early and late distant recurrence risk in breast cancer at the time of diagnosis.
• These are not incremental improvements. They are category shifts in what "early detection" means.

What Happened

On May 4, a team at the Mayo Clinic led by radiologist Dr. Ajit Goenka published a study in Gut, the BMJ's gastroenterology journal, that does something no human radiologist can do: spot pancreatic cancer on a CT scan more than a year before it becomes clinically visible.

The model is called REDMOD — Radiomics-based Early Detection Model. It was trained and validated on CT scans of patients who had been imaged for other reasons and later developed pancreatic cancer. The headline numbers:

• 73% sensitivity — it caught nearly three-quarters of cases that would otherwise have been missed.
• Median 475 days before clinical diagnosis — the model flagged abnormalities more than 15 months before a doctor would have found the tumour.
• In some cases, the lead time stretched to three years.

The study is retrospective. That's the limitation — and we'll get to it. But the finding itself is not subtle. This is not a marginal improvement on existing screening. Pancreatic cancer has no population screening programme because no test has ever been good enough. REDMOD, if prospectively validated, would be the first.

What It Actually Means

The mechanism matters. Pancreatic cancer is lethal not because it's untreatable in principle — early-stage tumours can be surgically removed — but because it's almost never found early. By the time a patient has symptoms (jaundice, weight loss, abdominal pain), the tumour has typically metastasised. The five-year survival rate is about 8% in the UK and similar across developed countries.

Goenka put it plainly to NBC: "We knew, based on the biology of the disease, that this is not something which is coming all of a sudden in three months."

The AI found what the biology predicted should be there. Specifically, REDMOD detected abnormal pancreatic cells that protect the disease from immune defences — a feature that human radiologists have never been able to see on CT. The model is not just doing what a radiologist does, faster. It is seeing something a radiologist cannot see.

That distinction — seeing the previously invisible rather than automating the visible — is what separates AI diagnostics that are incremental from those that are transformational.

The Two Breakthroughs, Side by Side

The CNN segment that aired on May 6 featured two AI cancer detection advances. The second, from Caris Life Sciences, is less visually dramatic but arguably more immediately deployable.

Caris MI Clarity, launched May 5, is the first AI-powered prognostic test that predicts both early and late distant recurrence risk in breast cancer at the time of diagnosis. It targets postmenopausal patients with HR-positive/HER2-negative, node-negative early-stage breast cancer — the most common subtype.

The clinical problem it solves is real: current tests predict early recurrence risk reasonably well, but late recurrence (5+ years after diagnosis) has been a blind spot. Patients who are told they're "low risk" based on early-recurrence models sometimes develop metastases years later. Caris MI Clarity closes that gap by training on both time horizons simultaneously.

Taken together, the two studies represent a convergence: AI is moving from assisting diagnosis to redefining what can be diagnosed, and from estimating risk to stratifying it across time horizons that previous tools couldn't reach.

The Limitations (Named Honestly)

Three limitations matter and none of them sink the finding:

1. Retrospective design. REDMOD was tested on scans from patients already known to have developed pancreatic cancer. The model knew what to look for because the outcome was already in the dataset. Prospective validation — running the model on current scans and following patients forward in time — is the only way to prove it works in the real world. The study's authors are explicit about this.

2. Sensitivity vs. specificity trade-off. 73% sensitivity means 27% of cancers were missed. That's better than zero (the current screening sensitivity), but it's not good enough for a standalone screening programme. The false-positive rate — how many healthy patients get flagged unnecessarily — hasn't been widely reported yet and will determine whether the model is clinically usable or merely promising.

3. Population diversity. The model hasn't been tested across ethnically diverse populations. Pancreatic cancer incidence varies by ethnicity, and AI models trained on one population can perform differently on another. This is a known problem in AI diagnostics and REDMOD hasn't addressed it yet.

None of these limitations are fatal. They are the normal constraints of a retrospective study. What makes REDMOD different from the dozens of other AI diagnostic papers published every month is the effect size. A median lead time of 475 days is not a rounding error. It's a clinical window.

The Pancreatic Cancer Problem, in One Paragraph

Pancreatic ductal adenocarcinoma (PDAC) accounts for about 95% of pancreatic cancers. It is the seventh-leading cause of cancer death worldwide and on track to become the second-leading cause by 2030 in Western countries. The reasons are structural: the pancreas is deep in the abdomen, tumours don't cause symptoms until they're large or metastatic, and there is no blood test or imaging protocol that reliably catches early-stage disease. The Mayo Clinic model doesn't solve all of these problems. But it solves the imaging problem — and the imaging problem has been the hardest one.

Stakeholder Landscape

Patients at risk — those with family history, genetic predisposition (BRCA2, PALB2, Lynch syndrome), or chronic pancreatitis — stand to gain the most. A model that can be run opportunistically on any CT scan taken for any reason transforms screening from a dedicated procedure into a background process.

Radiologists — the model doesn't replace them. It augments them. The most likely deployment pathway is triage: REDMOD flags suspicious scans, a radiologist reviews the flagged cases. This is the same model that has worked for mammography AI and lung cancer screening AI.

Health systems — the economics are compelling. Pancreatic cancer treatment is extraordinarily expensive (surgery, chemotherapy, palliative care). Catching it early enough for curative surgery saves money and lives. The barrier is not cost-effectiveness; it's the investment required to run prospective trials and integrate the model into existing CT workflows.

AI-in-medicine companies — REDMOD is a research model, not a commercial product. But the pathway from Gut publication to FDA clearance is well-established for AI diagnostics. Whoever licenses or replicates this approach will have a substantial market.

What This Means for You

If you are a patient with risk factors for pancreatic cancer: this model is not available to you yet. It hasn't been prospectively validated and it isn't deployed in any clinical setting. What you can do now: ensure any CT scan you receive for any reason is stored and accessible. If and when retrospective AI screening becomes available, having your scan history will matter.

If you are a clinician: the study is published in Gut (DOI: search "REDMOD Mayo Clinic Gut 2026"). Read the full paper, not the press release. Pay attention to the false-positive rate when it's published. Consider whether your institution has the data infrastructure to participate in prospective validation trials — Mayo Clinic will need partner sites.

If you are a health system administrator: the question is not whether AI cancer screening will arrive. It's whether your CT workflow and data architecture are ready for it. The model runs on routine CT scans. If your scans are siloed in departmental PACS systems with no API access, you're building technical debt that will delay deployment.

If you are a general reader: the honest answer is there is nothing actionable today. But the knowledge itself — that pancreatic cancer, one of medicine's hardest problems, may have met its match — is worth carrying. This is what progress looks like when it's done properly: a validated model, published in a peer-reviewed journal, with limitations named honestly, moving toward prospective trials.

Uncertainty Ledger

• Prospective validation: the single most important unknown. A prospective trial is the only way to confirm the 475-day lead time holds in the real world.
• False-positive rate: not yet widely reported. If it's high, the model may be useful for research but not for clinical deployment.
• Generalizability: will the model perform as well on non-Mayo Clinic populations, on different CT scanner models, on patients with different risk profiles?
• Regulatory pathway: FDA clearance for AI diagnostics typically requires prospective evidence. How long that takes depends on whether the FDA treats pancreatic cancer screening as a high-urgency indication.
• Caris MI Clarity: launched commercially, but real-world performance data won't be available for years given the 5+ year recurrence horizon it's designed to predict.

Bottom Line

The Mayo Clinic's REDMOD model detected pancreatic cancer on routine CT scans at a median of 475 days before clinical diagnosis — a lead time that, if prospectively validated, would transform the prognosis of one of medicine's most lethal cancers. The study is retrospective and the limitations are real, but the effect size is too large to ignore. Prospective trials are now a clinical imperative. Separately, Caris Life Sciences' AI-powered breast cancer recurrence test closes a blind spot that has troubled oncologists for decades — predicting late recurrence at the time of diagnosis. These are not incremental improvements. They are the first wave of AI diagnostics that don't just assist human judgment but extend it into territory human judgment cannot reach.

Sources

• Primary source — the Gut (BMJ) paper with DOI, lead author, model name (REDMOD), and key findings
• Tier 1 coverage — CNN, Mayo Clinic News Network, BMJ/Gut, The Ringer
• Tier 2 coverage — NY Post, The Mirror, Fierce Biotech
• Contextual data — American Cancer Society, NCI SEER, Cancer Research UK statistics on pancreatic cancer survival
• Related study — Caris Life Sciences breast cancer recurrence test (launched 5 May)
• Background — Topol's Deep Medicine, Lancet Digital Health, FDA AI/ML device clearances