The Med-AI Brief
Good morning, {{first_name|Doctor}}. Last issue: two scribes, one trial, two different answers. This week, the same problem at scale. Someone counted every AI device the FDA has ever cleared, then checked how many had been tested on patients.
In today's Med-AI Brief:
🩺 Research: three of 1,357 devices have outcome data
⚖️ Regulation: the FDA turns to generative AI
🌍 Use cases: Harrogate scans for HER2
💼 Jobs: eight roles at the companies building this
RESEARCH
🩺 1,357 cleared, three with outcomes
The brief: An analysis published in PLOS Digital Health on 19 August went through every AI-based medical device the FDA has authorised. Of 1,357, three had been tested on whether patients actually do better.
The details:
All devices authorised as of 5 December 2025. Just 34, or 2.5%, were linked to a registered prospective trial. Twelve posted results, twelve reached peer-reviewed publication.
Three, or 0.2%, evaluated a patient-centred outcome such as mortality, morbidity or readmission.
Where trials existed, they excluded pregnant women, adults over 75 and non-English speakers, and clustered in well-resourced centres.
The authors' conclusion: "Regulatory approval has outpaced clinical validation, creating an ecosystem where innovation advances without accountability."
Why it matters: Authorisation is not evidence of benefit. It is evidence that a device does what its manufacturer claims it does, usually measured against a predicate on retrospective data. That distinction survives about ten seconds into a procurement meeting. The count is FDA, but these are the same products that reach us carrying a UKCA or CE mark, backed by the same studies. So when a vendor says "FDA cleared", the useful follow-up is: cleared on which endpoint, and in whom?
REGULATION
⚖️ Both regulators started drafting
The brief: On 24 August the FDA said it is writing rules for generative AI in medical devices. In the UK, the National Commission's recommendations are due before the summer runs out.
The details:
Rick Abramson, who directs the FDA's Digital Health Center of Excellence, said to expect "broad guidance on the overall topic of generative AI" plus narrower specialty guidance. No date was attached.
The UK's National Commission into the Regulation of AI in Healthcare is chaired by Professor Alastair Denniston, with Patient Safety Commissioner Professor Henrietta Hughes as deputy. It was due to report "later this summer".
Its evidence base ran to 761 responses. Half wanted substantial revision of the current framework, 21% wanted a complete overhaul, and 65% wanted post-market surveillance improved.
Why it matters: Generative tools reached clinical workflows well ahead of any framework written for them, which is why the status of your scribe needed a separate clarification in July rather than being obvious from the rules. Two regulators are now drafting at once, and the drafts will decide what counts as evidence, who monitors a model after deployment, and where liability sits when it drifts. The moment clinicians get heard is the consultation. Consultations close quietly.
USE CASES
🌍 Harrogate scans for HER2
The brief: Harrogate and District became the first NHS trust to put an AI-assisted HER2 tissue scanner into routine use, announced on 18 August.
The details:
The Roche system captures high-resolution digital images of breast cancer tissue and supports pathologists assessing HER2 status, the marker that decides whether targeted therapy is an option.
It sits in the histopathology department at Harrogate District Hospital, with the trust citing faster diagnosis and quicker treatment decisions.
The £40,000 cost was met in full by the Harrogate Hospital and Community Charity, raised during its 30th birthday campaign.
Why it matters: Two things worth noticing. This is diagnostic AI, not admin AI, so unlike your scribe it sits squarely inside device regulation and someone in the trust owns that assurance. And a charity paid for it. When the capital route for clinical AI runs through fundraising rather than the capital programme, adoption follows which trusts have an active charity rather than which have the greatest need. Worth knowing how yours got funded before the business case for the next one lands.
JOBS
💼 Eight Opportunities, posted in the Health Tech Space
This week the vendors are doing the hiring, including three of the scribe companies from recent issues. All live as of this morning. NHS and academic posts return next week.
UK and Europe
Staff Machine Learning Engineer, Accurx, London. £135,000 to £155,000 plus share options up to £50,000. Apply
Implementation Lead, Secondary Care, Accurx, London. £70,000 to £90,000 plus share options. The one on this list a clinician could walk into. Apply
Customer Success Associate, Clinical Operations, TORTUS, London. Competitive salary and early-stage equity, hybrid. A UK ambient voice vendor. Apply
Senior Backend Engineer, Agents and Speech Recognition, Corti, Copenhagen. Apply
Senior Full-Stack Engineer, Nabla, Paris. The scribe that moved the needle in last week's trial. Apply
US and remote
Hiring in medical AI? Reply and I'll consider it for next week.
QUICK HITS
📰 Everything else
Screening: GRAIL's Galleri blood test goes before an FDA advisory panel on 23 September, the first multi-cancer early detection test to get one. Worth remembering that the NHS-Galleri trial, 142,000 people aged 50 to 77, missed its primary endpoint on late-stage diagnoses. Source
Gulf: If you are evaluating a tool for a Saudi facility, the SFDA already has binding AI-specific device requirements in MDS-G010, plus MDS-G27 covering digital health products. Ahead of the UK on paper. Source
UK: The MHRA's AI Airlock sandbox finished phase two with seven products and has £1.2m a year from DHSC through to 2029. Its findings feed the Commission above. Source
Tools: Hundreds of thousands of US doctors now use clinical LLMs from OpenEvidence, Doximity and UpToDate, sold as safer than general chatbots. The benchmarks behind that claim are contested. Source
How was today's brief?
Until next week,
Saeed
Know a colleague who'd use this? Send them your referral link.
Nothing here is clinical or legal advice. Check any tool against your own information-governance rules before it touches patient data.