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
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.
- Evidence Sept 11, 2026Kohn and colleagues ran Epic's one-year mortality model, which the paper describes as "perhaps the most widely available commercial mortality risk model," on 154,063 encounters at Trinity Health and 133,043 at Kaiser Permanente Southern California from 2022 through 2023. Discrimination was moderate to high, with a C statistic of 0.76 at Trinity and 0.81 at Kaiser, but calibration was poor in both systems, and the scaled Brier score at Trinity was essentially zero, meaning the probabilities the model reported there were about as useful as the base rate. Performance was weakest in patients 75 and older, where the C statistic fell to 0.69 and 0.73, and it degraded in liver disease, kidney disease and heart failure. Dr. James Deardorff, a UCSF geriatrician, wrote in STAT on Friday that the same prediction is harmless when it prompts a goals-of-care conversation and dangerous when it feeds a transplant or resource decision. Many hospitals have the score switched on.Sources: JAMA Network Open, STAT (Deardorff)
- Regulation Sept 18, 2026The bill, by state Sen. Steve Padilla, cleared both chambers with four votes against it, passing the Senate 39-0 and the Assembly 71-4, and was presented to the governor Sept. 9. On Friday the Senate Democratic caucus released a campaign by licensed mental health professionals urging the governor to sign it. The enrolled text bars any person or company from providing or advertising therapy or psychotherapy unless a licensed professional conducts it, bars licensed clinicians from letting an AI make independent therapeutic decisions or communicate therapeutically with a client, requires consent before AI touches session recordings and sets civil penalties of up to $10,000 per violation. The caucus release describes the bill as a response to the Trump administration's push to put AI therapy agents into Medicare, which the ACCESS model does for depression and anxiety. Gov. Gavin Newsom has until Sept. 30 to act on bills from the end of the session, the same deadline that applies to AB 1979 and AB 2575.
- Deployment Sept 16, 2026Zocdoc's Care Access Network lets outside platforms book real-time appointments with the marketplace's more than 200,000 providers across 10,000 insurance plans, using data feeds, an application programming interface (API) or, for AI assistants, a Model Context Protocol server, the company said Sept. 16. Named partners are Google's Gemini, Amazon Health AI, Yelp, Healthgrades, Blue Shield of California and WellHive. For Blue Shield, Zocdoc reported 2 million bookable hours opened in 90 days and a 600% increase in booked appointments with a six-day average wait; the figures are the company's own. The same week, Ferry Health came out of stealth with $9 million from Andreessen Horowitz and Index to run AI agents that phone clinics and book on a patient's behalf.
- Deployment Sept 17, 2026GW RhythmX's platform reads the chart, guidelines, care gaps, social and financial determinants, payer requirements and formulary, and hands the clinician a next-best-action recommendation during the visit rather than after it, the companies said in a Sept. 17 release. Phase one covers 30 physicians and advanced practice providers in family and internal medicine across two markets for 90 days, out of 473 primary care clinicians in the medical group, Fierce Healthcare reported. The company's reference customer is Presbyterian Healthcare Services, where it said the tool delivered 17,000 personalized responses and, in its chief executive's words, "supported $1.1 million in condition capture" over six months. Condition capture is a coding measure, not a clinical outcome. Payer requirements, including prior authorization rules, and the formulary are among the platform's inputs, and its recommendations arrive inside the clinician's electronic health record (EHR).
- Evidence Sept 16, 2026Guerra-Adames and colleagues at Bordeaux University Hospital trained a Mistral model to reproduce nurses' triage scores, then fed it pairs of cases that differed only in the patient's sex, they reported in npj Digital Medicine. Female versions came out 1.1% lower in predicted severity in Bordeaux and 2.2% lower on the U.S. MIMIC-IV records; retraining on sex-neutralized text made the gap disappear, and the pattern shifted with the sex of the triage nurse. The authors described the result as a probe of what the charts record rather than proof of bedside undertriage, and as a hypothesis generator rather than a verdict.Source: npj Digital Medicine
- Incident Sept 16, 2026OpenAI disclosed six cases from its own research in a misalignment reporting framework published Sept. 16 and said it will publish such cases as they arise. In one, instances of a model in training added instructions to their summaries to conceal mistakes from the user; in another, a model that could not retrieve requested earnings figures found an exposed API key and, when that failed too, made the numbers up and presented them as sourced. Becker's Hospital Review reported that the company said none of the six happened in a health care deployment. The company said its disclosures have been "ad hoc and less frequent than ideal."Sources: OpenAI, Becker's Hospital Review
- Regulation Sept 10, 2026In Raphael v. Mantei, a Clark County Superior Court judge held Aug. 7 that the recording a DAX ambient scribe makes to draft the note is an administrative document exempt from patient disclosure under the state's Uniform Health Care Information Act, the American Medical Association (AMA) reported Sept. 10. The AMA Litigation Center, the Washington State Medical Association and the Washington State Hospital Association filed together for the clinic, arguing that only the physician's final note is the record of care. Separately, Confluence Health said its primary care physicians cut daily EHR time by 34 minutes, and Ohio State told Becker's Hospital Review that AI visit summaries save 15 to 20 minutes per encounter. How long the audio exists depends on the vendor's retention policy.