The Week the FDA Let AI Treat Patients — Jun 26, 2026

0:00The FDA just cleared an AI that tells patients how much insulin to take. Welcome to MedAI Times Podcast, your daily update on medical AI. Don't forget to like and subscribe. Here's what moved in medical AI this week.

0:13The FDA cleared the first patient-facing medical language model, an AI that adjusts insulin for people with type 2 diabetes. An ECG-reading AI that just earned its own clearance is credited with triggering the world's first AI-attributed heart transplant.

0:28The FDA also cleared software that flags clogged leg arteries on routine CT scans. GE Healthcare spent $2.3 billion buying an imaging software company. GPT-5 Pro helped crack a three-year immunology puzzle.

0:44And Anthropic's most powerful model is still switched off, two weeks after Washington pulled the plug. Monday, we watched AI agents out-diagnose doctors on paper. This week, the FDA cleared one to actually treat patients.

0:59Start with the big one. A company called Updoc became the first to win FDA clearance for a patient-facing large language model that delivers care, not paperwork. The cleared job is specific. Adults with type 2 diabetes talk to the AI by voice or text, and it gives them insulin dosing instructions and adjusts medication between visits, wired directly into the electronic

1:22record under a licensed clinician's oversight. The clearance itself, a 510K, came through late last December. The company went public with it on June 25, alongside $18 million in seed money from Mayo Clinic, Eli Lilly, and Polaris.

1:37It's founded by Stanford Medicine researchers and already running at Cleveland Clinic, Allegheny Health Network, and UCSF Health. Why it matters. Almost every cleared medical AI to date flags or drafts for a doctor.

1:51This one speaks to the patient and changes a dose. So signal with an asterisk. It's a real clearance for a narrow task. Insulin titration, not the all-knowing AI doctor the coverage implies. Watch the safety data as the indication widens.

2:06Staying with the FDA. Pathway Labs, a spine-out from New York Presbyterian in Columbia, took its Echonex system public this week. It reads a standard 12-lead ECG and flags hidden structural heart disease across six indications.

2:21The peer-reviewed evidence in Nature put it at 77% accuracy versus 64% for cardiologists working without it. The new hook is a case report in Nature Medicine, the first heart transplant attributable to AI detection, after Echonex caught failure that clinicians had missed.

2:40Pathway is now pushing it through open evidence, the search tool used by a large share of U.S. clinicians. Another clearance, quieter but telling. Guide AI Health cleared a tool called Vascular Assist that reads routine CT scans and flags peripheral artery disease in the legs, the kind of blockage that often goes unnoticed until someone loses a limb.

3:02The company reported around 95% sensitivity at spotting affected patients. It's a triage and notification device, so it prioritizes scans for a radiologist rather than diagnosing on its own. Three FDA clearances in one week.

3:18All reading images are talking to patients. The pipeline is not slowing down. The money is moving too. GE Healthcare closed its $2.3 billion purchase of Intelerad, a cloud imaging software company doing about $270 million in revenue.

3:34This is consolidation, not a breakthrough. The read is that imaging AI has reached the boring lucrative phase. The value is shifting from clever detection algorithms toward the cloud plumbing that moves scans and reports between hospitals.

3:49An NVIDIA industry survey out this week says the same thing in different words, with most medical technology firms now reporting actual return on AI in imaging rather than pilots. The era of proving it has started.

4:03On the research side, OpenAI says GPT-5 Pro helped an immunologist solve a puzzle that sat unsolved since 2022. Daria Unutmaz, a 30-year veteran of T-cell biology at the Jackson Laboratory, fed the model raw, unpublished bench data.

4:20It proposed a mechanism his team had not tested, pointing at disrupted glycosylation and memory T-cells. The honest framing is his, not the marketing department's. The model generates hypotheses. Humans still design and validate the experiments.

4:34This is a vendor blog, single lab, not a peer-reviewed result. But it's a clean example of a frontier model as a lab collaborator on cold data. And the story that won't resolve. Anthropic's frontier models, Fable 5 and Mythos 5, are still completely offline, 13 days after a U.S. export control order tied to fears the weights could reach Chinese or

4:57Russian intelligence. Anthropic confirmed on June 25 that traffic is zero, viral claims that the models were backtraced to a user interface bug. Fable 5 also quietly left the free subscription tiers on June 23.

5:12The biomedical labs that had been using Mythos for molecular design work are still locked out. Spotlight. The thread running through this week is the patient-facing clinical language model as a regulated medical device.

5:27When an AI drafts a note, a human reads it before anything happens. When an AI tells a patient to take six units of insulin, the model is the last step before action. That moves the regulatory question from accuracy on a benchmark to safety in a kitchen at 7 a.m. and the FDA just said yes to the first one.

5:46All the links, sources, and the papers behind each story are in the description. We'll see you Monday. Thanks for listening. Find us on YouTube and your favorite podcast app. See you tomorrow.