Regulation
LLM device gets FDA breakthrough status
Startup RecovryAI came out of stealth today with the announcement of an FDA breakthrough designation for its patient-facing chatbot focused on recovery from joint replacement surgery. Large language models challenge the FDA's longstanding approach to device validation, and the agency has yet to authorize a device that relies on the technology.
But RecovryAI's submission, projected for the third quarter this year, could offer hints about the FDA's approach to regulating LLM-based medical devices.
Read more about how the company is determining its device's performance thresholds with the FDA in Katie's latest.
Artificial intelligence
Epic, Abridge, and others offer AI wish list to HHS
In December, the federal health department issued a request for information about how to boost the use of AI for clinical use. There's a ton to unpack in the responses, including proposals about reimbursement, modifying privacy regulations, and more.
In a first pass at the comments, I identified proposals from high-profile companies I thought were interesting or unusual, including ideas from Abridge, Aidoc, Doctronic, Epic, Oracle, and Tempus AI.
Read more here
policy
Claude gets the boot from HHS
In keeping with President Trump's broader directive, the Department of Health and Human Services is discontinuing use of Anthropic's large language model Claude. An enterprise version of Claude has been available to staff department wide. The AI model also underlies Elsa, an FDA tool launched last year that can be used by staff reviewing products.
HHS staff were told to instead use the department's installations of OpenAI's ChatGPT or Google's Gemini models. It's a bit more complicated for Elsa, which was built on Claude. However, Gemini is already available through the tool and will become the primary model going forward.
The federal government is ending its work with Anthropic following a dispute over how the company's AI tools might be used by the Department of War.
Medical devices
RadNet acquires AI company Gleamer for $250 million
STAT's Katie Palmer writes: RadNet, which owns more than 400 radiology practices in the U.S., continued its acquisition spree on Monday by buying up French radiology AI company Gleamer. Along with other radiology AI companies, RadNet has been racing to build and sell suites of algorithms that detect, diagnose, and triage findings across the entire body. Gleamer brings four U.S. regulatory clearances to the party.
RadNet will pay approximately $250 million when the deal closes and could pay $17 million more, dependent on hitting certain milestones.
The ultimate goal — though one without a clear regulatory precedent in the U.S. — is to automatically draft comprehensive radiology reports from medical images like CT scans and X-rays. Gleamer sells a tool that automatically converts AI findings into draft reports in Europe. "We're at the beginning of this," said Sham Sokka, chief operating and technology officer of digital health at RadNet, in a press conference. "With Gleamer now, we really want to move that journey toward much more of an automated radiology context over the next three to five years."
In its earnings release yesterday, RadNet reported just over $92 million in digital health revenue for 2025 and expects the segment to grow between 46% and 56% in 2026.
No comments