Microsoft has delivered the strongest peer-reviewed evidence yet for AI-based diagnosis. On October 22, the company announced that research on MAI-DxO (Microsoft AI Diagnostic Orchestrator) has been published in Science: on a simulated benchmark of complex cases drawn from the New England Journal of Medicine, MAI-DxO achieved 83% diagnostic accuracy, while practicing general practitioners and internists with access to reference materials scored 36%.
1. How the orchestrator works: simulating a multidisciplinary consultation
MAI-DxO is not a bigger single model. A unified orchestrator coordinates multiple AI agents that emulate an expert panel: they interview the patient sequentially, order tests, and debate differential diagnoses — narrowing candidates the way a clinical team would. Microsoft ran the architecture on foundation models from OpenAI, Google, Anthropic, Meta and DeepSeek; OpenAI's GPT-5 and GPT-5.5 performed best. The gains come from orchestration multiplied by base-model capability, not either alone.
2. Two benchmarks: doubled accuracy, half the cost
Key results
① NEJM complex cases: MAI-DxO 83% vs 36% for generalists (specialists, who normally treat such cases in multidisciplinary teams, were not in the comparison);
② MIMIC-IV-ED emergency admissions: MAI-DxO 80.3% vs 46.8% — a narrower ~1.7x edge;
③ Cost: reportedly half to one-third of a physician's, reaching the right diagnosis with fewer tests and procedures.
The narrower margin on real-world emergency data clarifies the current positioning: not triage replacement, but a high-quality “second opinion” for difficult cases.
3. What stands between paper and clinic
Three gaps remain: ① the comparison baseline — 83% vs 36% is against generalists, not the specialist teams such cases actually reach; ② liability and regulation — malpractice attribution and medical-device review pathways for AI diagnosis are still being defined; ③ access — multi-model inference costs less than a physician but requires infrastructure and data-compliance maturity, so large institutions will adopt first.
Healthcare combines extreme error costs with extreme knowledge density. A multi-agent consultation architecture validated here is a directly transferable pattern for law and finance.
For China's industry, another signal: the orchestration layer is a near-level playing field — DeepSeek models are already in Microsoft's test matrix. NineZenith's Tianxing LLM centers on the same capability: industry agent orchestration with private deployment and full audit trails, so an “AI second opinion” runs inside an institution's own controlled environment.
(Compiled from Microsoft's announcement, IT Home, Cailian Press and Lianhe Zaobao.)