THE MED-AI BRIEF
October 27, 2026 · 5-MINUTE READ · Read online
Good morning, {{first_name|Doctor}}. By now someone in your WhatsApp group has posted it: on Thursday Science published a Microsoft study in which an AI out-diagnosed generalist doctors 83% to 36%. Sceptics will call it a stunt. It isn't, quite. The number worth repeating is a different one, and it's in the third bullet below.
In today's Med-AI Brief:
🩺 Research: Microsoft's AI beat generalists. Twice.
🌍 Use cases: An AI scribe that listens in Arabic, live at seven Gulf sites
⚖️ Regulation: NICE's yes to five prostate AI tools comes with a clock
🛠️ Practice: four questions before you forward "AI beats doctors"
RESEARCH
🩺 Microsoft's AI beat generalists. Twice.
The brief: On 22 October Science published Microsoft's study of MAI-DxO, a system that runs AI models as a simulated panel of doctors: starting from the presenting complaint, it asks questions, orders tests and argues its way to a diagnosis.
The details:
Two text-only test sets: 304 NEJM case records, chosen for teaching value, and 325 routine emergency admissions from a US database.
On the NEJM cases the system scored 83.2%. Generalist physicians (primary care, internal and emergency medicine), working alone with search and references, scored 36.1%; 22.3% without them.
On the routine emergency cases: 80.3% against 46.8%. That is the fairer comparison, and the gap is still wide.
Across the models tested, the panel method added an average of 7.9 points on NEJM cases and 4.4 on emergency cases. The rest is the underlying model.
Every author is a Microsoft employee with equity in the company. Costs are estimates from one US hospital's price list. A real-world evaluation at Beth Israel Deaconess in Boston is next.
Why it matters: The emergency result is harder to wave away than the NEJM one. But the authors list the limits themselves: text only, no imaging, no patients in the room, individual doctors rather than teams, and diagnosis rather than outcomes. They expect the gap to narrow against specialists. Read it as strong evidence that frontier models are good at structured diagnostic reasoning on paper, not that they are ready for your clinic. The Boston study is the one to wait for.
USE CASES
🌍 An AI scribe that listens in Arabic, live at seven Gulf sites
The brief: On 22 October, at Abu Dhabi's Future Health Summit, M42 launched Ossi, an AI platform that sits alongside the electronic record rather than replacing it.
The details:
Four agents: a pre-visit patient summary, an ambient scribe that drafts a structured note for the doctor to review, a chat that answers questions about the record and checks against guidelines, and a coding assistant.
It listens in Arabic and English. It is live in seven centres in the UAE and Bahrain, including Healthpoint, Mubadala Health Dubai and Imperial College London Diabetes Centre.
M42 estimates doctors could spend "up to 60 per cent more time with patients". No method or study is given for that figure.
Cloud-hosted. No health regulator approval is mentioned.
Why it matters: This is the ambient scribe model NHS trusts are piloting, built for Arabic and already running across seven sites. The number that matters for Gulf clinicians is not the 60%; it is how accurately Ossi captures consultations that switch between Arabic and English, and M42 has not published that yet. "At the end of the day, who will make the final decision is the doctor," says M42's Kareem Shahin. Which is also where the liability sits.
REGULATION
⚖️ NICE's yes to five prostate AI tools comes with a clock
The brief: On 23 October NICE published draft guidance recommending five AI tools that help radiologists read prostate MRI, for NHS use during a two-year evidence-generation period.
The details:
The five: AI-Rad Companion Prostate MR, DeepHealth Prostate, ProstatID, Prostate Intelligence and QP-Prostate. Two others, hProstate and mdProstate, were not recommended.
The tools highlight areas that may contain cancer. A radiologist still reviews the scan and the AI output before any decision.
Each tool needs appropriate regulatory approval, including NHS England's DTAC (its baseline digital safety and data checks).
Sites must gather real-world data on accuracy, biopsy numbers, waiting times, radiologist workload and patient outcomes. The draft is open for comment.
Why it matters: This is NICE saying "use it, and prove it", not a final endorsement. Prof Prabhakar Rajan put the gap plainly: it is important to distinguish "promising performance" from "evidence of improved patient outcomes". If your department adopts one of these, you are part of the evidence. For Gulf buyers who cite NICE in procurement, a conditional recommendation should be read as exactly that.
PRACTICE
🛠️ Two minutes: four questions before you forward "AI beats doctors"
The Microsoft story will be passed round as "AI twice as good as doctors". Four questions sort a result from a headline:
Beat whom? Generalists, specialists or a team? This week: generalists, working alone.
On what? Real patients, or written-up cases in a simulation? This week: text-only cases.
Measured how? The right diagnosis on paper, or better outcomes for patients?
Run by whom? A developer's own study is a starting point, not a verdict. This week: every author works for Microsoft.
If the press release doesn't answer them, the methods section will.
QUICK HITS
📰 Everything else
Riyadh, from today: the Global Health Exhibition runs 26 to 29 October under the theme "Invest in Health", with more than 130,000 healthcare professionals expected. We will report what the SFDA and Ministry of Health announce.
Qatar: on 20 October the Amir ratified the establishment of a National Center for AI. It is a general AI body, not a health one; we will watch whether health falls within its remit.
Watching: MHRA guidance on updating AI devices after approval (December); the draft implementation plan for the AI Commission's 44 recommendations (end of 2026); comments on NICE's prostate MRI draft.
Know a colleague who has already forwarded you the Microsoft headline? Forward them this. They can subscribe at themedaibrief.com.
Until next week,
Saeed