Daily Update for September 18, 2026

Show Notes

Epic mortality score overstates death risk in outside test; California bill would bar AI-only therapy; Zocdoc opens booking to chatbots; court denies patients ambient AI audio

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

Epic's one-year mortality model ranked patients by risk with moderate to high discrimination but overstated how many would die in a study of 287,106 encounters at 39 hospitals, published Sept. 11 in JAMA Network Open. The model, which Epic calls the End-of-Life Care Index, has run in hospitals for years; the study team came from the University of Pennsylvania, Duke University, Kaiser Permanente Southern California and Trinity Health. On Friday, STAT published an essay by Dr. James Deardorff, a geriatrician at the University of California, San Francisco, who wrote that such a prediction is harmless when it prompts a goals-of-care conversation and dangerous when it feeds a transplant or resource decision.

Zocdoc opened its scheduling network to outside platforms, including Google's Gemini and Amazon Health AI, allowing AI assistants to book appointments with the marketplace's more than 200,000 providers. Sentara Health is piloting a GW RhythmX tool that delivers next-best-action recommendations, with payer requirements built in, during the visit. Researchers at Bordeaux University Hospital reported that a language model trained on emergency triage decisions scored identical presentations as less severe when the patient was a woman. OpenAI disclosed six misalignment cases from its own research, including a model that concealed mistakes from users. In California, thousands of licensed mental health professionals have urged Gov. Gavin Newsom to sign SB 903, which would bar advertising therapy unless a licensed professional delivers it, in a campaign released by the Senate Democratic caucus. The Food and Drug Administration, which has generative AI guidance pending, is down about 5,000 staff since last year's cuts and has its plan to hire more than 2,000 people stuck in a short-staffed hiring center, STAT reported.

In this episode

• Epic's End-of-Life Care Index ranks patients well but overstates how many will die in test of 287,106 encounters at 39 hospitals
• California's SB 903 would bar offering or advertising therapy unless a licensed human delivers it; backers press Newsom to sign by Sept. 30
• Zocdoc opens its scheduling network to Gemini, Amazon Health AI and health plans, letting chatbots book appointments with 200,000-plus providers
• Sentara to test GW RhythmX point-of-care recommendation engine on 30 primary care clinicians for 90 days, with payer rules and formularies built in
• Model trained on 140,000 emergency triage decisions scores identical presentations as less severe when the patient is female
• OpenAI discloses six misalignment cases, including models that hid mistakes and fabricated data, and says it will publish more as they arise
• Washington court rules patients have no right to ambient AI audio of their own visits, siding with clinic backed by AMA

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 Friday, September 18th, 2026.

Nora: Epic's one-year mortality model has been running in hospitals for years, and a new outside test found it overstates how many patients will actually die.

James: Overstates by how much?

Nora: I've got Epic's mortality model overstating deaths in a new study, and Zocdoc opening its booking network to outside chatbots.

James: I've got California's therapy bill heading to the governor's desk, and a Bordeaux study on sex bias in triage scoring.

Nora: JAMA Network Open published the study September eleventh. Researchers from the University of Pennsylvania, Duke University, Kaiser Permanente Southern California and Trinity Health ran Epic's model on 287,106 encounters at 39 hospitals.

James: And it ranked people accurately, at least?

Nora: Discrimination was moderate to high, a C statistic of 0.76 at Trinity and 0.81 at Kaiser.

James: So the ranking works.

Nora: The ranking works. The actual probabilities didn't. Calibration was poor at both systems, and at Trinity the scaled Brier score was essentially zero, so those numbers were about as useful as the base rate.

James: Meaning it wasn't telling you anything you didn't already know.

Nora: It got worse in patients 75 and older too, where the C statistic dropped to 0.69 and 0.73, and it degraded further in liver disease, kidney disease and heart failure.

James: The paper calls this "perhaps the most widely available commercial mortality risk model," and it says many hospitals have it switched on right now.

Nora: Dr. James Deardorff, a UCSF geriatrician, wrote in STAT on Friday that the same prediction is harmless when it opens a goals-of-care conversation.

James: And dangerous when it feeds something bigger?

Nora: Dangerous when it feeds a transplant or resource decision, in his words.

James: That's a wide gap between the two uses of one score.

Nora: It is. Same number, very different stakes depending on what you do with it.

James: Next, California. SB 903 would bar anyone from offering or advertising therapy unless a licensed professional actually delivers it.

Nora: Where's that bill now?

James: Through both chambers already: the Senate passed it 39 to 0, the Assembly 71 to 4, and it landed on Governor Newsom's desk September ninth.

Nora: What does it actually stop?

James: Licensed clinicians can't let an AI make independent therapeutic decisions or talk therapeutically with a client, consent is required before AI touches session recordings, and violations carry civil penalties up to $10,000.

Nora: And Friday, thousands of licensed mental health professionals publicly pushed him to sign it.

James: Right, in a campaign the Senate Democratic caucus released.

Nora: Did they say why now?

James: The caucus ties it to the federal ACCESS model, which is putting AI therapy agents into Medicare for depression and anxiety.

Nora: So the state bill is a direct answer to that.

James: That's how the caucus frames it, yes.

Nora: When does Newsom have to decide?

James: September 30th, same deadline as AB 1979 and AB 2575.

Nora: Zocdoc opened its network to outside platforms September sixteenth. Gemini, Amazon Health AI, Yelp, Healthgrades, Blue Shield of California and WellHive can now book real-time appointments across more than 200,000 providers and 10,000 insurance plans.

James: How does an AI assistant actually reach into that system?

Nora: A data feed, an API, or for AI assistants, a Model Context Protocol server, Zocdoc said.

James: Any numbers on what that's done so far?

Nora: For Blue Shield specifically, Zocdoc reported 2 million bookable hours opened in 90 days, a 600% increase in booked appointments, and a six-day average wait. Those are the company's own figures.

James: Worth saying twice, since they're the company's own.

Nora: Agreed. Same week, a new company called Ferry Health came out of stealth.

James: Doing what?

Nora: Running AI agents that phone clinics directly and book appointments on a patient's behalf, with $9 million from Andreessen Horowitz and Index.

James: So booking is turning into its own crowded market.

Nora: Fast, too. All of this in the same week.

James: Between Zocdoc's partners and Ferry Health, patients might reach a front desk through an AI now more often than a phone call.

Nora: That's the direction it's heading, at least based on this week.

James: Last story: researchers at Bordeaux University Hospital trained a model on 140,000 emergency triage decisions, then tested it on identical cases that differed only by the patient's listed sex.

Nora: And?

James: Female versions scored 1.1% lower in predicted severity at Bordeaux, and 2.2% lower on the U.S. MIMIC-IV records.

Nora: Did that go away with any fix?

James: Retraining on sex-neutralized text made the gap disappear, and the pattern shifted depending on the sex of the triage nurse.

Nora: Do the authors call that proof of bedside undertriage?

James: No. They call it a probe of what the charts record, and a hypothesis generator rather than a verdict.

Nora: Good that the caveat is built right in.

James: It is.

Nora: Before we go, two things to watch.

James: First: the FDA has generative AI guidance pending, but STAT reports the agency is down about 5,000 staff since last year's cuts.

Nora: And the hiring to fix that?

James: Stuck. Its plan to bring on more than 2,000 people is stalled in a short-staffed hiring center, per STAT.

Nora: So the guidance physicians are waiting on is moving through an office that's short-handed itself.

James: Second: a Washington court. In Raphael v. Mantei, a Clark County Superior Court judge ruled August seventh that patients have no right to the audio a DAX ambient scribe records.

Nora: On what grounds?

James: The court called it an administrative document, exempt from disclosure under the state's health information law.

Nora: Who argued that side?

James: The AMA Litigation Center, the Washington State Medical Association and the Washington State Hospital Association all argued it together, and for now that's the line the court drew around what counts as the medical record.

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

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