Nature tests MIRA and AMIE as clinical workflow agents
Nature published two papers testing medical AI agents in clinical work: MIRA achieved 87.8% diagnostic accuracy versus 78.1% for four board-certified physicians on 311 ICU cases from MIMIC-IV, and AMIE matched or exceeded physicians on outpatient management; both validated in simulation only.
On 17 June, Nature published two papers where AI agents did clinical work: taking histories, ordering tests, revising diagnoses, building treatment plans. In a head-to-head comparison on 311 cases from MIMIC-IV (a de-identified database of real ICU patient records), MIRA reached the correct diagnosis in 87.8% of cases; four board-certified physicians scored 78.1%.
MIRA operated inside an EHR sandbox (a virtual electronic health record), ordering labs and imaging, prescribing medications, and proposing admissions. The system ran on 574 cases total. Of 468 prescriptions, 467 were rated clinically correct; the main errors involved route of administration (intravenous vs. oral vs. subcutaneous).
Google DeepMind and Google Research expanded AMIE for longitudinal disease management: 100 clinical cases, three visits with a standardized patient, compared against 21 primary care physicians. The system drew on UK NICE guidelines, BMJ Best Practice, and drug references. In blinded specialist evaluation, AMIE matched physicians on clinical reasoning and exceeded them on prescription accuracy, test ordering, and guideline adherence.
Both results come from simulations. MIRA never saw a real patient. AMIE worked with standardized patients in a virtual scenario. Real clinical practice adds queues, rescheduled appointments, forgotten tests, conflicting complaints, and economic pressure on decisions. Neither paper claims prospective trials in real clinical settings.
Eternal Search tracks Google as an organization.
Open the related Eternal Search page
Sources
t.me
nature.com
nature.com
AMIE is a Google DeepMind product, and Eternal Search already tracks Google as an organization with a health, AI, and longevity funding profile, making this Nature paper connectable to a broader investment trajectory rather than an isolated research event.