AI Completes the Full Emergency Care Workflow: MIRA Outperforms Physicians in Simulated Diagnostic Accuracy, but Remains Far from Clinical Deployment

Emergency care decision-making is not a one-time answer, but a sequence of choices that alter one another: taking a medical history, ordering tests, revising the diagnosis, and then deciding on medication, surgery, or hospitalization. A study published in *Nature* shows that a medical AI agent called MIRA can complete this entire workflow within a simulated electronic health record. This comes closer to clinical practice than simply answering case-based questions, but it still cannot be equated with caring for real patients.

Teams including the Dresden University of Technology and University Hospital Dresden created an isolated testing environment using 574 existing emergency department cases from the MIMIC-IV database, covering eight diseases including appendicitis, pancreatitis, pneumonia, pulmonary embolism, urinary tract infection, and pancreatic cancer. Through 11 tools, MIRA could perform more than 85,000 possible actions, including taking medical histories; ordering and interpreting blood, microbiological, and imaging tests; proposing differential diagnoses; and arranging medications, procedures, and hospitalization. Every action was converted into a structured electronic health record entry.

Across all cases, MIRA achieved an average diagnostic accuracy of 88.9%. The researchers separately selected 311 cases to compare the system with four specialists using the same interface and under the same information conditions. MIRA achieved an accuracy of 87.8%, compared with 78.1% for the physician group. Another group with mixed levels of experience, comprising four resident physicians and two specialists, achieved 71.1%. The advantage was not uniform: MIRA performed particularly well on appendicitis and pancreatitis, where test results were more definitive, but both MIRA and the physicians had lower accuracy for pneumonia and urinary tract infection.

MIRA also demonstrated an ability to translate diagnoses into actual medical orders, such as identifying the need for surgery, reconciling pre-admission medications, and recommending hospitalization. The researchers spot-checked 468 prescriptions for 56 patients. Of these, 467 had correct written dosing instructions rated as relevant and clinically useful; however, accuracy was lower for the route of administration, at 97%. These data support the possibility of using AI as a workflow assistant, but also serve as a reminder that high overall scores may still conceal a small number of errors that matter to individual patients.

Independent experts noted that MIRA ordered about twice as many blood tests as the physicians. Obtaining more information could itself improve diagnostic accuracy, so “the same conditions” does not mean that decision-making costs and resource use were entirely equivalent. The study also compared the system with only a small number of physicians; interactions were limited to text and a maximum of 20 rounds of dialogue; the cases were curated retrospective data; and only eight diagnoses were covered. The system did not encounter patients’ ambiguous descriptions, disruptions in clinical settings, or complex conditions it had never seen before.

The findings are therefore better viewed as an engineering and validation milestone for medical AI agents than as clinical proof of an automated emergency physician. The next step will require prospective studies across different hospitals and patient populations, while also clarifying methods of supervision, resource use, error interception, data governance, and the allocation of responsibility. The research team positions MIRA as a collaborative assistant to physicians, with ultimate clinical responsibility still to be borne by physicians.