TU Dresden on-prem medical AI hits ~90%, defers uncertain cases

A fully on-premises clinical AI agent from Prof. Jakob N. Kather's group at TU Dresden diagnosed patient cases at roughly 90% accuracy on one benchmark and 84% on a second, publishing in Nature Medicine on 15 September.

The system runs as two agents talking to each other: one plays the physician, one plays the patient, working from anonymized electronic records including laboratory values, medications and examination findings. The conditions in scope were appendicitis, cholecystitis, pneumonia, pulmonary embolism and urinary tract infections. On 181 randomly selected cases, physician reviewers agreed with the automated evaluation in more than 90% of cases.

The more interesting result is what the paper says about knowing when to trust the model. Repeating the diagnosis several times and measuring how stable it stayed turned out to be the strongest predictor of correctness; the model's own internal probability scores were unreliable. The framework routes stable calls through and defers unstable ones back to a clinician.

"Our goal is an AI agent with selective autonomy," Kather said in a statement released by EurekAlert. "These systems should support clinicians in decision-making, but never take over completely." The system builds on MIRA, the group's earlier autonomous medical AI agent from June 2026, with collaborators at Dresden University Hospital and the National Center for Tumor Diseases in Heidelberg.

The paper also notes weaker diagnostic performance on simulated cases involving older patients. Two researchers we track surfaced the paper on publication day.