Microsoft reports MAI-DxO reached 80% diagnostic accuracy versus physicians’ 47% on simulated MIMIC-IV-ED cases in a Science study; prospective hospital testing is still planned.

Microsoft’s SDBench diagram explains the sequential diagnostic evaluation: an agent requests tests or asks questions, a gatekeeper supplies case information, and a judge scores the diagnosis. The depicted 304-case branch is the NEJM dataset, separate from MIMIC-IV-ED.

Published 22 Oct 2026 · 19:41 UTC

Microsoft AI ↗ · Source dated 22 Oct 2026

Microsoft reports MAI-DxO reached 80% diagnostic accuracy versus physicians’ 47% on simulated MIMIC-IV-ED cases in a Science study; prospective hospital testing is still planned. · ournews.ai