Physician in the Loop

AI flags esophageal cancer on plain CT; Rush hands chart abstraction to a vendor; AI firms sponsor MAHA Summit; Newsom vetoes prior authorization bill

Show Notes

AI flags esophageal cancer on plain CT; Rush hands chart abstraction to a vendor; AI firms sponsor MAHA Summit; Newsom vetoes prior authorization bill

The voices on this episode are generated, and so is the script, from the day's post at Physician in the Loop. Every fact in the post links to its source.

Read the post and its sources

A Chinese research group reported Tuesday in Nature Medicine that a model reading the esophagus on noncontrast CT outperformed all 17 radiologists in a reader study and was validated on 80,612 patients. The model flags esophageal cancer and high-grade precancer on the plain scan ordered for kidney stones and lung nodules. The validation covered 12 centers in three countries, and the model was then run prospectively on hospital scans and on 10,959 consecutive people in a low-dose lung screening program. The paper appeared one day after Alibaba released an abdominal CT model free on GitHub.

The Food and Drug Administration (FDA) press page lists nothing newer than a gene therapy approval on Thursday, and the Centers for Medicare & Medicaid Services (CMS) on Monday announced preliminary lab fee schedule rates. President Donald Trump on Friday announced an AI Force and a coming AI czar with no detail beyond a promise not to hinder the industry and a statement that the existing criminal and civil courts can handle bad behavior, Axios and Fortune reported; the announcement so far includes nothing on medicine. The Make America Healthy Again (MAHA) Summit on Sept. 29 is sponsored by OpenAI, Anthropic, Hims & Hers, Noom, Sword Health, Tempus, Prenuvo and Centene, among others, and sponsors receive dinner with the speakers, according to the summit's site and Becker's Hospital Review. In Sacramento, Gov. Gavin Newsom on Sunday vetoed a prior authorization bill that had passed 64-4 and 31-0 and has not acted on four health-AI bills, with eight days left to do so.

Elsewhere, a survey of 13 countries found that 49% of respondents believe a person with no medical training but skilled with AI could do at least one physician task as well as a physician; DexCare bought Mila Health in the second front-door agent acquisition in a week; and Becker's Hospital Review reported on how health systems triage the AI features Epic ships by the dozen, in a piece in which Rush's chief information officer said on the record that a chart-review job went to a model and is not coming back.

In this episode

What is Physician in the Loop?

How technological innovation is changing medicine for patients and the people who practice on the front lines.

Nora: This is Physician in the Loop, the daily briefing on AI in medicine. I'm Nora Davis.

James: And I'm James Hudson. It's Tuesday, September 22nd, 2026.

Nora: A model reading plain CT scans, the kind ordered for kidney stones or a lung nodule, just outperformed all 17 radiologists at spotting esophageal cancer.

James: On scans nobody was even looking at the esophagus for. That's the part that gets me.

Nora: Exactly. Nature Medicine published it Tuesday, and it was validated on over 80,000 patients across three countries. I'm leading with that esophageal CT model, and then we'll get into Rush handing a chart-abstraction job to a vendor for good.

James: I've got the MAHA Summit sponsor list, and Newsom's veto of a prior authorization bill in Sacramento.

Nora: So this model, called EAGLE, comes from a Chinese research group. Two stages: one finds the esophagus on the scan, the other segments any lesion and scores it for malignancy.

James: And it actually held up prospectively, not just in testing?

Nora: It did. On an external set of 11,466 patients it hit 90% sensitivity and 98.5% specificity for cancer, and 52.5% sensitivity for high-grade precancer. Run forward on real hospital scans and on 10,959 people in a lung screening program, its flags had a positive predictive value of 42.2%, with specificity of 99.94% in that screening group.

James: In the reader study, how much did it beat the radiologists by?

Nora: It beat all 17 of them, by an average of 19.1 points of sensitivity and 18.4 points of specificity. And when it assisted the radiologists instead of replacing them, their own cancer sensitivity jumped from 71.9% to 85.7%.

James: There has to be a catch given how it was built.

Nora: The authors say it themselves. Nearly all the patients were Chinese, where squamous cell cancer dominates, and validation abroad was only preliminary. It also performed worse in women because the training set skewed male, and follow-up was under two years. In the U.S. the common tumor is adenocarcinoma of the distal esophagus, which this model hasn't been shown to catch.

James: So a real result, with a very specific population attached to it.

Nora: That's the honest read. And it landed one day after Alibaba put an abdominal CT model up free on GitHub.

James: Speaking of jobs changing shape: Becker's Hospital Review talked to seven health systems about the AI features Epic now ships by the dozen with every upgrade.

Nora: And Rush's story stood out.

James: Rush's chief information officer, Jeff Gautney, said a clinical registry task that used to take five people reviewing charts, under an outsourced contract of about $1 million a year, now runs on a vendor called Layer Health.

Nora: What happened to Epic's own version of that feature?

James: Gautney said Epic's been "threatening for two years to deliver a product." About those five jobs, he was blunt: "That's not a job that's coming back."

Nora: Did other systems sound as certain?

James: Not quite. Seattle Children's runs an AI Review Board that ranks features by expected return before piloting anything, because, as its chief medical information officer put it, you can't run a six-month pilot on every single one. Michigan Medicine's former CIO described the more common approach: start with Epic, and even if it isn't as good, it's good enough.

Nora: And Brooklyn Hospital Center wasn't in a rush to test any of it.

James: Their line was: "We prefer not to go first." Now to the MAHA Summit on September 29. Its own site lists the sponsors: OpenAI, Anthropic, Hims and Hers, Noom, Sword Health, Tempus, Prenuvo and Centene, among others.

Nora: That's a lot of AI and telehealth money in one room. What do sponsors actually get?

James: Dinner with the speakers, according to the summit's site and Becker's Hospital Review. Those speakers include Anthropic's Syed Mohiuddin, OpenAI's Felipe Millon, and Dr. Jay Bhattacharya, who directs the National Institutes of Health.

Nora: Any numbers on what that access costs?

James: A report relaying Bloomberg's account put the top sponsorship at $300,000, for a speaking slot, input on panel topics, and a dinner for 30 with Robert F. Kennedy Jr. and Dr. Mehmet Oz. That figure hasn't been independently verified, and HHS says the department had no part in planning the sponsorships.

Nora: And no physician society made the sponsor list.

James: None. It's worth flagging the timing too: this comes a week before comments close on the FDA's generative-AI paper, and while Medicare officials are discussing a payment category for AI physicians.

Nora: Last stop before the watch list: Sacramento. Gov. Gavin Newsom's Sunday legislative update listed 83 bills signed and 24 vetoed.

James: One of those vetoes was a prior authorization bill?

Nora: AB 539, from Assemblymember Pilar Schiavo. It would have required health plans to honor a prior authorization for at least a year, or for the length of a shorter treatment. It passed the Senate 31 to 0 and the Assembly 64 to 4, and the California Hospital Association backed it.

James: That's a wide margin to veto.

Nora: It is, and to be clear, it's not an AI bill. Four actual health-AI bills, AB 1979, AB 2575, SB 903 and SB 503, weren't on Sunday's list or Monday's either, when he signed data-center bills on water and power but said nothing about health.

James: So those four are just sitting there.

Nora: With eight days left for him to act. One of the questions they'd settle is whether an AI can direct a medical assistant.

James: Before we go, two things to watch.

Nora: First, that Edelman Trust Barometer analysis, done with the Yale School of Public Health and reported by Becker's Hospital Review. Across 13 countries, 49% of people believe a person with no medical training but skilled with AI could do at least one physician task as well as a physician.

James: That's the belief, not a measured outcome.

Nora: Right, the survey measured what people think, not what actually happened. And it climbs to 59% among adults under 35.

James: Second: DexCare bought Mila Health, the second front-door agent acquisition in a week. Mila's agents call, text and chat with patients to book and prep their visits.

Nora: Any numbers behind that?

James: DexCare says the agents get a 54% response rate on outreach, cut no-shows in half and lift revenue 18%. Those are the company's own figures, with no published study behind them.

Nora: Worth remembering which numbers come from a company and which come from a journal.

James: Especially in the same week we're reading a Nature Medicine paper that spells out its own limitations line by line.

Nora: Sources and corrections are at physicianintheloop.org. I'm Nora Davis.

James: I'm James Hudson. See you tomorrow.