In detail
- German system MIRA (TUD Dresden, Heidelberg University, others) achieved correct diagnoses in 88.9% across 8 disease categories on 500+ MIMIC‑IV emergency cases.
- Head‑to‑head on 311 cases: MIRA 87.8% vs. four experienced specialists 78.1%; residents/specialist mix 71.1%.
- MIRA runs as an autonomous agent inside a sealed virtual EHR, selects from >85,000 options across 11 tools (labs, microbiology, imaging, treatment plans, admissions).
- Google's AMIE produced more accurate treatment and testing plans; both agents use base models that are already outdated.
Why it matters
Shows agentic, domain‑specific AI can deliver clinically relevant decisions in controlled settings — important for healthcare providers, vendors and regulators — but results are retrospective/simulated and depend on base models that evolve quickly.
For you Assess agentic AI through controlled pilots (triage, decision support) with strict monitoring and liability review before operational use.