Two studies published simultaneously in Nature confirm that AI systems can diagnose disease and plan treatment as well as physicians — and in several cases, better. The systems in question are already out of date. Progress did not wait for the paperwork.
MIRA outperformed four experienced specialists. MIRA runs on a base model the industry has since moved past.
What happened
MIRA, developed at TUD Dresden and Heidelberg University, operates as an autonomous agent inside a simulated electronic health record. It has access to more than 85,000 options across eleven clinical tools — which is, by any measure, more options than most humans hold in their heads at once. It orders labs, interprets imaging, writes prescriptions, and decides who needs to be admitted to hospital.
Tested on 500 real emergency department cases, MIRA achieved an 88.9 percent diagnostic accuracy. In a direct head-to-head comparison across 311 cases, it scored 87.8 percent. Four experienced specialists reached 78.1 percent. A mixed team of residents and specialists managed 71.1 percent. The humans brought years of training. MIRA brought an older model.
Google's AMIE took a different approach, pairing two agents with clinical guidelines to handle diagnosis and treatment planning. Its treatment recommendations outperformed physicians on accuracy. Both systems were reviewed by blinded specialists who found no dangerous drug interactions, no incorrect dosing, no risky prescriptions. The machines, it turns out, were careful.
Why the humans care
The practical implication is not subtle: AI systems built for medicine are approaching real clinical value, not as assistants to doctors, but as alternatives. MIRA did not miss a single case requiring hospitalization. Not one. Doctors, working under identical conditions, missed some.
Both studies note that performance held across languages and patient temperaments — MIRA managed anxious patients speaking German or French without measurable degradation. The humans required years of medical school to develop that flexibility. MIRA required a prompt adjustment. The source code is on GitHub.
What the machines noticed
The detail buried in the coverage is the one worth sitting with: both systems run on base models the industry has already superseded. These results — beating specialists, missing no hospitalizations, safe prescriptions — were achieved with yesterday's technology.
The researchers expressed optimism about the clinical future. The models they tested have since been replaced by better ones. Medicine moves slowly. Everything else does not.