Emergencies rarely happen in isolation. In such high-stress situations, doctors must simultaneously ask questions, review lab results, evaluate images, and make decisions. What could help: an AI assistant that takes over routine tasks for them. Researchers at Dresden University of Technology and Dresden University Hospital have now developed just such a system. They call it MIRA.
MIRA stands for “Medical Intelligence for Reasoning and Action.” The system is what’s known as an AI agent—a program that independently plans and executes tasks. Unlike many previous AI applications, MIRA operates directly within a simulated electronic health record.
In a simulated emergency room, the system processed more than 500 real patient cases. It asked questions, ordered tests, evaluated findings, and prepared treatment recommendations. MIRA not only reached the level of a physician but even outperformed the physician control group in diagnostic accuracy in some areas. The study was published in the journal *Nature*.
How MIRA Works in Practice
The research team led by Prof. Jakob N. Kather from the Else Kröner Fresenius Center for Digital Health at TU Dresden set clear boundaries for MIRA. Eleven tools were available, including lab orders, imaging, and medication prescriptions. This allowed the agent to choose from more than 85,000 possible tests, treatments, and decisions. MIRA also interacted with virtual patients whose responses were based on real medical records.
“Our AI agent was able to independently perform clinical procedures within the test environment. MIRA identified missing information, requested tests, interpreted findings in accordance with guidelines, and prepared treatment decisions,” says Dyke Ferber, a physician and the study’s first author. "In accordance with guidelines" means that the system adhered to medical recommendations and treatment standards.
Responsibility Remains with Humans
Despite the positive results, the researchers emphasize that MIRA is not a substitute for doctors. “I see AI agents as autopilots in an airplane. Such systems can support and relieve medical professionals by taking over routine tasks, but the responsibility ultimately always remains with the medical staff,” says Prof. Kather.
Prof. Uwe Platzbecker, Chief Medical Officer at Dresden University Hospital, sees great potential in the results. The study shows that AI agents can support diagnostic decisions and prepare treatments. The next step is to integrate such systems safely and transparently into clinical practice.
Original publication:
D Ferber, L Hilgers, C Höper, B Kinny-Köster5, JN Eckardt, K Egger-Heidrich, M Bill, MMK Schneider, J Clusmann, L Kadric, M Oehme, M Mayrhofer-Schmid, A Oeser, G Wölflein, IC Wiest, JM Middeke, AJ Iafrate, D Truhn, D Jäger, JN Kather. Towards Autonomous Medical Artificial Intelligence Agents, Nature 2026.
