The Miracle Diagnostic That Might Not Save Who You Think

Yesterday, the Mayo Clinic published a landmark study in the journal Gut showing that an AI model called REDMOD can detect pancreatic cancer on routine CT scans up to three years before clinical diagnosis¹. The tool identified 73% of cancers that human radiologists had originally read as normal. For scans taken more than two years before diagnosis, REDMOD was three times more accurate than specialist radiologists². This is, by any honest measure, extraordinary science.

Pancreatic cancer kills more than 85% of the people it touches, mostly because we find it too late. By the time symptoms appear, the disease has usually metastasized. The five-year survival rate is 13% overall but jumps to 44% for localized disease⁴. That 31-point gap is the entire prize. REDMOD is designed to reach into that gap and pull people back from the abyss. The clinical team is now advancing a prospective trial, AI-PACED³, to evaluate real-world performance. If the results hold, this could redefine pancreatic cancer from a death sentence into something closer to a treatable disease.

I want to celebrate that. I also think the press coverage so far is missing something critical.

Here's what bothers me. A 2025 National Cancer Database study of 75,801 patients⁵ found that Black race, lower socioeconomic status, and uninsured status were all associated with "stable lower odds" of receiving surgery or chemotherapy for pancreatic cancer across the entire period from 2010 to 2017. The word "stable" is doing a lot of work in that sentence. It means the disparities did not narrow at all over nearly a decade of clinical advancement. Separately, a widely cited review in PMC⁶ reported that uninsured patients had adjusted odds of receiving surgery at just 0.07 compared to insured patients. That's not a small disadvantage. That is functionally no access to curative treatment.

Think about what this means for an AI that finds cancer three years early. If you're a privately insured patient getting an abdominal CT at the Mayo Clinic for some unrelated complaint, REDMOD is potentially life-saving. It catches a shadow on your pancreas that a radiologist would have missed, you get fast-tracked into surgery, and you join the 44% five-year survival cohort. That's a genuine miracle. But if you're an uninsured patient in rural Mississippi, or a Medicaid recipient at a low-volume community hospital, what does an early detection flag actually get you? The ASCO GI 2025 presentation on persisting disparities⁷ reported that Black patients are still less likely to be referred to surgeons and, once referred, still less likely to receive resection. The pipeline from detection to treatment is broken at every joint.

I want to be fair about the counterargument, because the counterargument is strong. The survival gradient for pancreatic cancer is so steep that even a modestly functioning system that catches more cancers at Stage I will save some lives that would otherwise be lost. And there's a real historical pattern in which clinical breakthroughs create political pressure for access reform. The Massachusetts insurance expansion in 2006⁸ was associated with a 67% increase in pancreatic cancer surgical resection rates for previously underserved patients. That's a powerful proof of concept: when you give people coverage, they get surgery, and they live.

But the Massachusetts story actually proves my concern rather than refuting it. Coverage came first, then resection rates rose. The diagnostic tools and surgical techniques already existed. What was missing was the insurance card. And when we look at the broader Medicaid expansion data, the picture is less encouraging. A March 2026 JAMA Surgery study of 51,707 patients⁹ found that while Medicaid expansion was associated with higher surgical resection rates and lower 2-year mortality, the improvements were "delayed and uneven" and crucially did not reduce income-related disparities. A University of Pittsburgh analysis¹⁰ found that Medicaid expansion produced earlier-stage diagnoses but "no differences in rates of surgery, postoperative outcomes, or overall survival." This is the part that should make everyone uncomfortable. Even when we expand coverage, the treatment gap for the poorest patients persists for pancreatic cancer specifically.

The TB diagnostics story offers a parallel that cuts both ways. In March 2026, the WHO recommended a new class of battery-powered TB tests costing half the price of GeneXpert, delivering results in under an hour with minimal infrastructure. That is genuine progress driven by sustained upstream innovation. But GeneXpert itself, recommended by the WHO since 2010, is still not equitably deployed after 14 years. MSF's "Time for $5" campaign¹² has documented that Cepheid, owned by Danaher, continues to charge $15-20 per test for XDR-TB, HIV, and hepatitis in low-income countries despite production costs of roughly $3-5. A coalition of over 150 organizations delivered 206,937 petition signatures¹³ to Danaher's headquarters in October 2024 demanding price reductions. As of the last reporting, Danaher had reduced the primary TB test from $10 to $8 but left most other tests at monopoly pricing. Innovation created the tool. Political economy decided who gets to use it.

I think the honest position here is uncomfortable for both camps. The people who say "build it and the system will follow" are ignoring the 70-year mammography experiment (Black women still face 40% higher breast cancer mortality¹⁴ despite similar screening rates) and the 14-year GeneXpert deployment stall. The people who say "don't build it until the system is ready" are proposing something that has never happened in the history of medicine and that would condemn patients to Stage IV diagnoses while waiting for a political consensus that may never arrive.

But if I have to pick, I think the weight of evidence favors building the tool while treating access reform as an absolute precondition for deployment, not an afterthought. The key insight from the peer-reviewed literature on AI and cancer equity is clear. A 2025 review in Current Oncology Reports¹⁵ found that AI in cancer care "risks exacerbating existing health disparities" when access to the technology "remains limited, particularly in low-income and rural settings." And the PMC review⁶ on pancreatic cancer disparities found something genuinely revealing: in equal-access systems like Kaiser Permanente and the U.S. Department of Defense, racial disparities in pancreatic cancer treatment and survival disappeared. The tool isn't the problem. The system is.

This means REDMOD's clinical trial design, reimbursement pathway, and regulatory approval process need to embed equity mechanisms from day one. That looks like: (1) mandating that the AI-PACED prospective trial recruit across diverse socioeconomic and racial populations, not just Mayo Clinic patients; (2) conditioning any eventual FDA clearance or CMS reimbursement on demonstrated equal performance across demographic subgroups; and (3) pairing diagnostic deployment with patient navigation programs and insurance coverage mandates, the way Germany has required AI-assisted detection software¹⁶ in its national lung cancer screening program.

My position, stated plainly: REDMOD is a tool that could save thousands of lives. Deployed into the current U.S. healthcare system without structural reform, it will save thousands of lives that belong disproportionately to wealthy, insured, urban patients. That outcome is better than saving no one. But it is not "changing everything," and pretending otherwise gives policymakers permission to avoid the harder work of expanding treatment access.

What to watch next: the AI-PACED prospective trial's enrollment demographics will be the first real signal. If the trial population mirrors Mayo's patient base (predominantly white, privately insured, high-income), the equity problem will be baked in before the first clinical result is published. The second indicator is whether CMS proposes a reimbursement code for AI-assisted pancreatic screening in the next 18 months. Without a payment mechanism, community hospitals and safety-net institutions won't adopt it regardless of the science. And the third thing to track: whether any of the 10 states that still haven't expanded Medicaid move to do so. Because the best diagnostic tool in the world is worthless if the patient it flags can't afford the surgery that follows.