Voice 1: This is Physician in the Loop, the daily briefing on AI in medicine. It's Wednesday, October 21st, 2026.
Voice 1: Researchers at Mass General Brigham say an AI model trained on 40 million face photographs sorted cancer patients into risk groups whose mortality differed more than threefold.
Voice 2: Threefold, just from a photo. How does that actually work?
Voice 1: The model, called FAHR-Face, was fine-tuned into two tools. One, FAHR-FaceAge, estimates biological age, with a mean error of about 5 years on public datasets. The other, FAHR-FaceSurvival, was trained on about 34,000 photographs of patients with cancer.
Voice 2: And that's where the threefold difference comes from?
Voice 1: Yes. Patients in the highest-risk quartile had more than triple the mortality of the lowest quartile, even after adjusting for clinical prognostic factors, and it held up in a second, independent cohort. The study ran in The Lancet Digital Health.
Voice 2: Do the two tools look at the same part of the face?
Voice 1: No, they draw on different facial regions, and combining them improved accuracy further.
Voice 2: So is this heading toward actual clinical use?
Voice 1: Not yet. Dr. Raymond Mak, a radiation oncologist and the corresponding author, said the aim is to support treatment decisions, “rather than replace physician judgment.” The health system has also filed a provisional patent application on facial health algorithms.
Voice 2: No prospective clinical use reported, in other words. Just the retrospective result for now.
Voice 1: The FDA had a different kind of announcement, about apps sold alongside prescription drugs.
Voice 2: What did the agency actually decide?
Voice 1: It finalized guidance on when software paired with a drug belongs in the drug's prescribing information. If the app is tied to a device used with the drug, like one paired with an inhaler, it generally should be described there.
Voice 2: What about standalone apps, the dose reminders and coaching tools?
Voice 1: Those only belong in the prescribing information if the drugmaker shows they're essential to safe and effective use, or deliver a meaningful clinical benefit. Otherwise, the content most companion apps show users will be treated as promotional labeling.
Voice 2: That's a real line to have to defend.
Voice 1: Jared Seehafer, the FDA's deputy commissioner for technology, said companion apps “are quickly moving beyond simple dose reminders toward tools that adapt to each patient.”
Voice 2: A smaller company made its own announcement this week, about phages instead of apps.
Voice 1: Radical Numerics says phages designed by its AI model, Omnii-1, were given to a patient with an antibiotic-resistant infection.
Voice 2: Did it actually work?
Voice 1: The company says the phage cocktail cleared the patient's resistant bacterial isolate in laboratory testing before it was given through a physician-led compassionate-use pathway. It calls this the first clinical use of AI-designed phages.
Voice 2: What aren't they telling us?
Voice 1: A lot. The company didn't name the bacterium, the hospital, or the treating physicians, and didn't disclose the patient's clinical course or outcome.
Voice 2: So we don't even know if the patient got better.
Voice 1: Right. The company itself says one patient can't establish safety or efficacy, that the therapies aren't approved for broader use, and that no paper on the case has been published.
Voice 2: A different kind of oncology story came out of Rhode Island this week, about keeping people out of the hospital rather than treating them once they're there.
Voice 1: What's the program?
Voice 2: Daymark Health's home and virtual supportive oncology program, launched with Blue Cross and Blue Shield of Rhode Island across 20 oncology practices. It was tied to about 10% fewer hospital admissions among more than 5,400 cancer patients.
Voice 1: And the cost side?
Voice 2: Spending came in about 8% lower than projected, or $789 per patient per month.
Voice 1: Projected against what, exactly?
Voice 2: Against trends built from two prior years of claims, not a randomized control group. The authors themselves say this shows an association, not causation, and the findings were presented as an ASCO abstract.
Voice 1: A real number, with a real asterisk attached to how it was measured.
Voice 2: Before we go, something to watch.
Voice 1: The HIMSS AI in Healthcare Forum opens in San Diego on October 22nd, running two days.
Voice 2: What's actually on the agenda?
Voice 1: Governance, policy, and implementation, the vendor and CIO view of what's being rolled out next.
Voice 2: Worth watching whether any of today's tools, the face model, the companion apps, the phages, come up as examples of what that governance is supposed to catch.
Voice 1: Sources and corrections are at physicianintheloop.org, where you can sign up for the free email briefing. This briefing was read by AI voices.