THE MED-AI BRIEF
October 20, 2026 · 5-MINUTE READ · Read online
Good morning, {{first_name|Doctor}}. On Thursday JAMA published the first large study of how clinicians actually use an AI tool at work: 270 million questions from 950,000 US prescribers. Most of them were about treatment. And many were asked outside the clinician's own specialty, which is exactly where a wrong answer is hardest to spot.
In today's Med-AI Brief:
🩺 Research: 950,000 clinicians, 270 million questions
🌍 Use cases: patients texted their EEG, and false alarms fell
⚖️ Regulation: the MHRA's plan to 2030 gets an AI centre
🛠️ Practice: check the source when you're out of your lane
RESEARCH
🩺 950,000 clinicians, 270 million questions
The brief: On 15 October JAMA published an analysis from Harvard Medical School and Brigham and Women's Hospital of how US physicians, physician assistants and nurse practitioners use OpenEvidence, an AI tool that answers clinical questions from the published literature.
The details:
950,000 clinicians, about two and a half years of use, 270 million queries. The team classified a random sample of 38.8 million with a language model.
51.9% of queries were about clinical decisions, mostly treatment and management. 13.9% were about diagnosis.
Monthly active users rose from 355 in January 2024 to 703,265 in June 2026.
Clinicians often asked about treatment outside their own field; surgeons, for example, asked plenty of non-surgical questions. Early-career doctors and oncology, haematology, internal medicine and family medicine asked more than their share.
Three co-authors work for OpenEvidence, which supplied the de-identified data but did not fund the study.
Why it matters: This measures use, not accuracy, and the authors say so: whether the answers are right, and what they do to patient care, is the next study. The finding to take away is where the questions land. "They are often using it in areas where they don't have specific expertise," says senior author Anupam Jena. That is the setting in which a fluent answer is most persuasive and least checkable. It is US data from one tool, but the behaviour will look familiar on any ward round where someone has a phone out.
USE CASES
🌍 Patients texted their EEG, and false alarms fell
The brief: On 13 October Nature Medicine published a Penn Medicine study in which inpatients on an epilepsy monitoring unit could ask an AI about their own live EEG, and tell the seizure detector when it was wrong.
The details:
13 adults at the Hospital of the University of Pennsylvania, several days each. More than 1,300 messages between patients and the AI.
Patients labelled more than 67% of real-time seizure detections as true, false or uncertain. Using those labels to update the detector cut false alarms by about 80%. Expert review added roughly 2% more.
The AI answered more than 92% of patients' questions about their brain activity accurately and relevantly.
"Completely unprompted, several patients asked if they could take the system home with them," says co-first author Zack Goldblum.
Why it matters: Thirteen patients in one unit is a proof of concept, and the authors call it that; they want larger outpatient studies next. The idea is the interesting part. Monitoring devices are usually built for the clinician, and the patient is a passive source of signal. Here the patient becomes a labeller, which is cheap, and the detector learns from it. Any monitoring service fighting alarm fatigue should watch where this goes.
REGULATION
⚖️ The MHRA's plan to 2030 gets an AI centre
The brief: On 14 October the MHRA published its strategy to 2030. AI runs through it, both in the products it regulates and in how the agency works.
The details:
A new MHRA Centre for AI and Digital Health, to set "clearer regulatory expectations" for developers of AI-driven devices and medicines. No launch date or budget yet.
"A modern, adaptive regulatory framework for AI as a medical device."
Expanded change control plans (agreed rules for updating a model after approval), and master files for foundation models (the large general models many tools are built on).
Real-time safety surveillance using UK data, plus an AI Medicines Development Action Plan in 2027.
Why it matters: A strategy sets direction rather than rules, so nothing changes in your trust tomorrow. But it confirms the thread from last week: regulators now treat an AI tool as something that changes over time, and plan to watch it after approval rather than sign it off once. On 19 October the MHRA also said it is recruiting a liaison officer to the FDA in Washington. For Gulf hospitals that buy on the strength of US or UK clearances, closer alignment between the two is worth watching.
PRACTICE
🛠️ Two minutes: check the source when you're out of your lane
The JAMA study suggests the AI questions clinicians ask most often sit outside their own specialty. Three habits for those moments:
Open the cited paper or guideline for any dose, drug interaction or threshold. If the answer has no citation, treat it as a colleague's opinion, not a reference.
Check that the source is current and fits your setting. A US guideline may not match NICE, or your local formulary in Riyadh or Dubai.
Keep the question free of patient identifiers. Age, key comorbidities and the drug are usually enough.
Thirty seconds each. When you can't judge the answer yourself, the citation is the only safety net you have.
QUICK HITS
📰 Everything else
Abu Dhabi: On 20 October, the opening day of the Department of Health's Future Health Summit, AstraZeneca and Oracle announced an AI partnership. Its first programme, with M42, will help teams find patients with uncontrolled hypertension, support guideline-directed care and recruit to trials. It is a non-binding term sheet; AstraZeneca says it will not receive patient-identifiable data.
Abu Dhabi, also day one: Johnson & Johnson said it is extending its collaboration with the Department of Health to standardise operating theatres on its Polyphonic digital platform. The summit runs to 22 October.
Next week: the Global Health Exhibition in Riyadh, 26 to 29 October. We will report what the SFDA and Ministry of Health announce.
Watching: MHRA change-control guidance (December); NHS England and NHS Resolution indemnity guidance for AI (no date); HSSIB's ambient voice investigation.
How was today's brief?
✉️ One favour: hit reply and tell me which AI tool you reach for when a question falls outside your specialty, and whether you open the reference.
Until next week,
Saeed