A.I. news · October 23, 2026
Microsoft MAI-DxO Will Be Tested on Real Hospital Patients
On complex NEJM cases MAI-DxO achieved 83% accuracy versus 22% and 36% for physicians without and with reference resources.
🔎 What Microsoft announced
Think of MAI-DxO as a panel of virtual physicians. Several A.I. agents question a patient, order tests and debate a diagnosis. Each one proposes diagnoses, picks useful questions or tests, and challenges the emerging view.[1][2]
The debate runs over several rounds. The system also weighs each test's value against its cost.[1]
📊 The numbers, and why a new test was built
On the complex cases, MAI-DxO got 83% right. Physicians got 22% without reference resources and 36% with them. On the typical emergency cases, it scored 80%, versus 47% for physicians.[1]
Tuned for cost, one configuration hit 81% accuracy at less than half the cost of comparative systems. Microsoft says no other approach was both more accurate and less costly across both data sets.[1]
Microsoft is open-sourcing SDBench for select NEJM cases. It also says the gains persisted and improved with newer models, including GPT-5.5.[1]
🏥 The hospital study
Microsoft has a new partnership with Beth Israel Deaconess Medical Center to evaluate clinical reasoning in hospital settings. In a prospective study, patients will interact with the system directly. It will be measured against physicians' independent assessments.[1]
Dominic King, vice president of Health at Microsoft AI, spoke to TechCrunch about the setup. Patients will get standard treatment. Those who consent will also engage with DxO in parallel, and its recommendations will be compared with the doctors'.[2]
King says the study will give compelling evidence by the end of the study period. He also says it will be the biggest study of its kind.[2]
⚠️ The fine print on those results
The 36% figure came from primary care doctors and general internists working individually with reference materials. Specialists or physician teams, who usually handle such complex cases, weren't part of the comparison.[2]
Microsoft tested models from several companies, with its best results from GPT-5 and GPT-5.5. GPT-5 alone scored 78.9% at about $8,400 per case. DxO cut the cost by about $3,000 and raised accuracy about four points. On emergency department cases, there was little improvement over GPT-5.[2]
With GPT-5.5, DxO raised accuracy on complex cases from 72% to nearly 87%. The gains didn't show up on emergency department cases. Gemini 2.5 Flash rose from 46.4% to nearly 69%, but the gap between models was larger than DxO's boost.[2]
Our take: the underlying model seems to matter most, and the panel approach looks like a useful add-on, especially on hard cases. Microsoft's own page stresses that gains persisted with newer models. The two readings may not tell the same story, so the full paper matters.[1]
🚀 The bigger race and Microsoft's plans
Microsoft joins Google and OpenAI in claiming its A.I. can match or outperform doctors. Google published a clinical study with Beth Israel in The Lancet. It tests AMIE in real-world urgent care visits, with physicians monitoring.[2]
Microsoft already sells dictation and radiology tools to clinicians and sees DxO as an expansion. It is bringing health teams, including those behind Dragon Copilot and PowerScribe, into an integrated effort at Microsoft AI.[2][1]
Suleyman said Microsoft must have the number one clinical decision support products over the next few years. He says Microsoft is working with the Mayo Clinic on a healthcare foundation model, planned for release next year. He predicts a coding-style breakthrough for medical records in 2027 and expects healthcare models to improve by an order of magnitude within 24 to 36 months.[2]
Those are forecasts from people building the thing, so hold them loosely. Our take: the race seems to be moving from exam scores to real clinics.
💡 What we don't know yet
King names the biggest challenge in healthcare delivery as a mismatch between demand and supply of medical expertise. If that's right, better tools could eventually help, but that's our hope, not a promise.[2]
We'd like to see the full Science paper explain the emergency department numbers. The thing to watch is whether the hard-case gains show up with real patients.
💡 Our take: what this could mean for you
- Judge any A.I. headline by its comparison group, the model underneath, and the data set used.
- Try a panel approach in your own work: have a model propose answers, challenge them, and weigh the cost of each next step. Track what a result costs, not only whether it's right.
- Treat health chatbots as a way to prepare questions for your clinician, and wait for the Beth Israel results before assuming benchmark gains carry over to real patients.
Have a business or a business idea? 🚀
Trinity builds a free, personalized website prototype for it — no credit card.
Get my free prototype