Microsoft diagnostic AI beats generalist physicians on hard cases, study says
Physician in the Loop, Oct 23, 2026
Listen to this post
read by AI voices ยท all episodesTHE SHORT READ
Microsoft's multi-agent diagnostic system reached the correct diagnosis in 83.2% of complex New England Journal of Medicine cases, against 36.1% for generalist physicians, in a benchmark study Science published online Oct. 22.
The physicians in the comparison were primary care physicians and general internists working alone with reference materials, and no specialists or physician teams took part, TechCrunch reported. The authors described the results as coming from a controlled benchmark, and Microsoft said it plans a prospective study at Beth Israel Deaconess Medical Center in Boston in which patients interact with the system directly.
Elsewhere, the director of the Food and Drug Administration's Digital Health Center of Excellence said Oct. 20 that draft guidance on generative AI devices will come in the coming months, MedTech Dive reported. Sutter Health said more than 6,000 of its clinicians have drafted more than 5 million notes with Abridge's ambient AI, and the FDA approved Edwards Lifesciences' Pascal Precision system for transcatheter repair of the tricuspid valve.
THE DETAILS
- FYIEvidence Oct 22, 2026The system, MAI-DxO, coordinates several AI agents that act like a panel of physicians: they question a simulated patient, order tests and revise a differential diagnosis while weighing test costs, according to the paper by 15 Microsoft authors, "Sequential diagnosis with language models." It reached 83.2% accuracy on New England Journal of Medicine clinicopathological conference cases and 80.3% on emergency department admissions from the MIMIC-IV-ED database, against 36.1% and 46.8% for generalist physicians; Microsoft said physicians working without reference resources scored 22% on the NEJM cases. Most of the improvement appeared to come from the underlying models, and on emergency department cases the orchestrator added little or no accuracy for GPT-5 or GPT-5.5, TechCrunch reported. A 2025 preprint of the work had reported 80% for the system against 20% for generalist physicians. Microsoft said a prospective study at Beth Israel Deaconess Medical Center will compare the system's performance with physicians' independent assessments; it gave no size or timeline.
- WatchRegulation Oct 21, 2026
Abramson said FDA draft guidance on generative AI devices will come in the coming months; he set no date.
Rick Abramson, director of the FDA's Digital Health Center of Excellence, said Oct. 20 at AdvaMed's MedTech Conference in Boston that the agency had received more than 500 submissions on its generative AI discussion paper, whose comment period closed Oct. 19, and that "we intend to draft policy, and it'll come soon," MedTech Dive reported. He called the policy "a priority at the highest levels of the administration," said a draft on generative AI devices for mental disorders is on the agency's top-tier list for fiscal 2027 and that an umbrella draft on risk assessment and evidence generation for generative AI devices is under construction. Dr. Michelle Tarver, director of the Center for Devices and Radiological Health, said at the same meeting that the center had about 22% attrition over the past year and a half, including up to 30% of its clinical reviewers, and has since hired more than 65 people, MedTech Dive reported. - FYIDeployment Oct 22, 2026Sutter Health says more than 6,000 clinicians have drafted 5 million notes with Abridge's ambient AISutter Health said its ambient documentation tool from Abridge has cut documentation time by up to 30%, and that average start-of-care documentation in home health fell from 97 to 55 minutes, Becker's Hospital Review reported from a Sutter release dated Oct. 21. Ten Aidoc radiology and stroke tools had analyzed more than 583,000 studies and flagged more than 43,000 cases for clinician review as of August, and about 158,000 patients had used the Emmie patient assistant as of September, the health system said. Of patients with stage 2 hypertension in its remote monitoring program, which had more than 14,000 enrollees, 90% reached blood pressure control within six months, according to Sutter. The figures are the health system's own and have not been published in a peer-reviewed study.Source: Becker's Hospital Review
- FYIEvidence Oct 21, 2026Authors at Google Research and Google DeepMind, with co-authors at Vituity and Columbia University, reported a multi-agent system they call a personal health agent: data science, health domain expert and health coach sub-agents under a Gemini-based orchestrator. This is a benchmark, expert and end-user rating study, not a clinical trial: no patient outcomes were measured. Among 45 end users, 46.7% ranked the full system first, against 28.9% for a parallel multi-agent baseline and 24.4% for PHIA, a single-agent system. Five clinicians reviewed the domain expert agent's summaries for 30 participants: 83% showed no or minimal potential harm, 15% mild, 2% moderate-to-severe. The authors said the work measured conversational quality and short-term task efficacy, not behavior change or longitudinal health outcomes, and that it is Fitbit-centric. The paper states that the study was funded by Alphabet and that most of its authors are current or former Alphabet employees who may own stock as part of standard compensation.Sources: Nature, Nature (PDF)
- FYIRegulation Oct 21, 2026The approval covers transcatheter edge-to-edge tricuspid valve repair in patients with symptomatic severe tricuspid regurgitation who are on optimal medical therapy, are at intermediate or greater surgical risk and are judged appropriate for the procedure by a multidisciplinary heart team, Edwards Lifesciences said Oct. 21. The stated aim is to improve quality of life and functional status. The release gave no outcome figures; one-year results from the CLASP IITR pivotal trial are to be presented in a late-breaking session at the TCT 2026 meeting, according to the company and TCTMD.Sources: Edwards Lifesciences, TCTMD
- WatchIncident Oct 22, 2026
Under the court's schedule, the FTC's opposition is due Nov. 25 and a hearing on the motion is set for Jan. 7, 2027.
The telehealth company filed its motion Oct. 22 in the U.S. District Court for the Northern District of California, where Judge Vince Chhabria is hearing the case, according to the court docket. The Federal Trade Commission, joined by California and Utah, sued July 29, alleging that Hims & Hers shared customer lists and information about users' activity on its sites with Meta, Snap and other advertising platforms through tracking tools, and enrolled consumers in recurring prescription subscriptions almost immediately after they submitted intake forms, MobiHealthNews reported. The company has called the suit baseless. The docket does not describe the grounds of the motion. - FYIMoney Oct 20, 2026FirstMark Capital led the Series A, with CVS Health Ventures and Valtruis joining, the company said Oct. 20. Baba said it has supported more than 10,000 patients and families, coordinating specialist visits, hospital discharges, behavioral health and needs such as housing and transportation, and that CMS selected it for the ACCESS model, Medicare's 10-year test of technology-supported chronic care. The company has published no patient outcome, cost or utilization data.Source: Baba (GlobeNewswire)