← News

Agents & MCPSources: arXiv · 19 Nov 2026

Decision2Care agent runs clinical work-ups without free text, logging every choice for audit

Researchers describe a decision agent that diagnoses acutely ill patients using probabilities and fixed rules instead of written reasoning, recording each option considered so clinicians can audit the work-up.

In plain words

AI agents that help doctors diagnose patients usually explain themselves in written text, but that text may not match what actually drove the decision. Decision2Care instead assigns a probability to each allowed action in the medical record system and follows fixed rules to act. Every option and probability is logged, so reviewers can see exactly why each step happened.

What was announced

The team behind Decision2Care evaluated the agent on records from 546 emergency admissions covering 12 diseases. Diagnostic accuracy reached 86.5%, close to the 86.1% reported for the MIRA-v2 agent, while the median work-up took 2.2 seconds against 92-100 seconds for the comparison agent. In 10 of the 12 diseases, the differential diagnosis shifted most at the test on which the relevant guideline bases the diagnosis, giving reviewers a traceable link between evidence and conclusions.

Confidence also signalled reliability: where the agent was most confident after a single work-up, all but four diagnoses were correct (98.5%), and the authors traced each of those four errors to a single imaging report, with three corrected once that report was removed. Treatment decisions used the same mechanism, and the agent's drug orders were judged guideline-adherent more often than MIRA-v2's. Its death prediction matched three intensivists on the same emergency cases, and in five of eight intensive care cohorts it exceeded the best of three 24-hour severity scores within 2-6 hours of records.

What it does not show

  • The evidence is an arXiv abstract, so evaluation details, peer review status and full methodology are not available from this excerpt.
  • Reported accuracy, speed and adherence figures come from the authors' own comparison against MIRA-v2 and have not been independently confirmed.
  • The abstract does not describe real-world deployment, regulatory review or availability of the system.

For security and IT teams

Scope
The published evidence covers retrospective evaluation on 546 emergency admissions across 12 diseases, not live clinical use.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.