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 Monday, September 21st, 2026. Nora: A model anyone can download for free just beat 23 of 26 radiologists reading abdominal CT scans. James: Free, no license fee, today? Nora: Today. Alibaba's DAMO Academy published it in Science on Thursday and put the code and weights on GitHub, and it still has no FDA clearance anywhere. James: Sounds like a story with some fine print. Nora: I've got that one, plus Oura's IPO filing and the physicians it's quietly routing members to. James: I've got a study on how stigmatizing language changes AI triage, and Heidi's new numbers on how many visits its scribe covers a week. Nora: So, the DAMO model. It reads a contrast-enhanced abdominal CT and checks for 146 findings across 18 organs, cancers included. James: That's a lot to screen in one pass. What did they train it on? Nora: More than 400,000 contrast-enhanced abdominal CT exams paired with their reports, and about 15 million anatomy-tagged image-text pairs, according to the GitHub page. James: And how did it actually perform? Nora: South China Morning Post reported a mean AUC of 0.913 across those 146 findings, tested on nearly 40,000 real exams from eight outside centers. It beat 23 of 26 radiologists head to head. James: What happens when a radiologist uses it as help instead of a rival? Nora: TechTimes reported about 10 points of sensitivity gained, and reads more than 30% faster. James: Could a hospital just run this tomorrow, then? Nora: The code is Apache 2.0, but the weights are licensed for research only, not commercial use. And it was validated on Chinese patients only, no prospective trial anywhere. James: And the Federal Register published a final order Thursday. Every radiology detection or triage algorithm still needs its own 510(k) clearance from the FDA, open weights or not. Nora: So download it, study it, but don't deploy it. James: Now the language study. Emergency physicians at UT Southwestern built patient pairs matched on Emergency Severity Index, one patient deteriorating within six hours, one not. Nora: And they ran that through more than one model? James: Gemma, Qwen and DeepSeek, across 221,556 comparisons. Stigmatizing language about frequent emergency department use pushed the deteriorating patient down the list, in every model, by as much as 9.4%. Nora: Did neutral wording fix that? James: Neutral phrasing scored 1.4 to 7.1 points higher, according to the paper in npj Digital Medicine. Nora: What about other attributes, psychiatric history, race? James: Psychiatric history moved rankings by up to 9.5%, but only reached significance in one model. Race, language and insurance showed no consistent harm. Nora: So the authors think the wording itself is the problem? James: They concluded the wording of the input is a safety-critical design choice. Nora: That tracks with Thursday's post, the model that scored identical presentations as less severe when the patient was female. James: Same theme, different attribute. How you describe a patient changes what the model decides. Nora: Oura launched its IPO Monday morning. 50 million shares, priced at $40 to $44, on the Nasdaq under the symbol OURA. James: Who's actually selling those shares? Nora: 13.5 million shares from the company itself, 36.5 million from existing holders. Goldman Sachs, Morgan Stanley and J.P. Morgan are leading the deal. James: What's that worth? Nora: At the midpoint, about $2.1 billion. That's less than the $3 billion deal at a valuation above $16 billion that Bloomberg reported in August, per Investing.com. James: What do the underlying numbers look like? Nora: The S-1 shows 5.0 million paid members as of June 30, revenue of $1.21 billion over nine months. Hardware brought in $974 million, membership $240 million, up 74%, and net income for the period was $60.8 million. James: Any word on who's wearing the ring? Nora: Fierce Healthcare reported the demographics: about 72% women, 27% over 45. James: Does the filing mention any of the medical features we've covered before? Nora: It describes an AI health companion called Oura Advisor and custom health LLMs, and calls the product a general wellness item it believes isn't subject to FDA device rules. James: But it doesn't mention Counsel Health. Nora: Not once. Even though Oura members have been able to reach Counsel's physicians inside the app since September 17, per Fierce Healthcare. James: Heidi's Asia-Pacific chief medical information officer, Darran Foo, told the HIMSS Asia-Pacific meeting the company's ambient scribe now supports 2.8 million patient visits a week, across 190 countries and 110 languages. Nora: Any hard result behind that number? James: One example: a New Zealand emergency department cut documentation time from 17 minutes to four per patient. Nora: Was that published anywhere? James: No. That's a vendor figure from a conference stage, unaudited, and no study was cited for it. Nora: What about the survey they released? James: Heidi surveyed 1,823 clinicians across 25 countries in July. 86% said they use AI daily or several times a week. Nora: Any policy behind that use? James: 83% said they do it with no employer guidance and no formal policy. 68% named hallucination as their top concern. Nora: That 83% lines up with the Imprivata survey we covered on September 15, where 72% of health system leaders reported AI deployed without IT approval. James: Different surveys, same gap. Clinicians are already using this every day, and most workplaces still haven't written the policy for it. Nora: Sources and corrections are at physicianintheloop.org. I'm Nora Davis. James: I'm James Hudson. See you tomorrow.