New AI Agent Outperforms Human Clinicians in Emergency Diagnosis, Study Published in Nature
A new artificial intelligence agent designed to manage the full workflow of emergency care has demonstrated diagnostic accuracy higher than that of experienced human clinicians, according to a pair of peer-reviewed papers published in Nature. Named MIRA, the AI system was built by an international research team to read electronic medical records, analyze lab results, assess patient symptoms, list differential diagnoses and propose treatment plans, rather than only completing single imaging recognition tasks.
The research team tested MIRA on thousands of real, de-identified emergency cases from multiple hospitals across North America and Europe. In blind comparison tests, the AI reached an overall diagnostic accuracy of 87.8%, while the panel of board-certified emergency physicians scored an average of 78.2%. The model also showed stronger performance in identifying easily-missed high-risk conditions, including aortic dissection, sepsis and rare neurological emergencies. Unlike earlier medical AI tools which usually focus on CT scans, X-rays or blood report analysis alone, MIRA works as an end-to-end clinical agent that can follow the whole logic of doctor-patient consultation.
Researchers explained that MIRA was trained on millions of complete clinical encounter records and learned to weigh conflicting signs, track the progression of symptoms and avoid common cognitive biases that often lead doctors to misdiagnosis. It can also explain its reasoning step by step: showing which lab value, symptom or past medical history supports its judgment, a feature known as explainable AI that is critical for clinical adoption.
However, the scientists repeatedly emphasized that the AI is not meant to replace doctors. It is positioned as a decision-support assistant, especially useful in hospitals facing shortages of emergency specialists, rural clinics and night-shift settings where medical resources are stretched thin. Many low- and middle-income regions lack enough experienced emergency physicians, and the team hopes this type of AI can reduce fatal delays before patients receive proper treatment.
The publication has triggered wide discussion across global medical and tech circles. Medical experts point out several big challenges ahead: data privacy risks, adaptability to different ethnic groups and local disease profiles, legal responsibility when AI gives wrong advice, and the risk of clinicians over-reliance on machine outputs. Regulators in the EU and the U.S. have launched review frameworks for clinical AI agents, requiring long-term real-world safety monitoring before formal approval for routine emergency use.
The lead researcher added that the next version of MIRA will be adjusted to fit different healthcare systems, and the team will carry out multi-country clinical trials to verify its stability in diverse medical environments. This milestone marks a shift of medical AI from auxiliary image reading toward independent reasoning in complex, high-stakes clinical decision-making.
