Five hospitals built one AI brain — without showing each other a single patient file

Imagine five hospitals that want to build one shared brain, but none of them is allowed to show the others a single patient file. For years the answer was: then no shared brain. This paper shows the workaround now works at clinical quality. Each hospital teaches a language model on its own records, locally. Only the tiny mathematical updates — never the data — travel to a central server that merges them into one model.

Across five health systems and 2.1 million de-identified patient notes, the shared model came within 1 point of a model trained the forbidden way, with all data pooled centrally. More telling: it beat the best model any single hospital could train alone by 12 points. They also stress-tested privacy — standard attacks designed to squeeze patient text back out of the updates recovered nothing once protective statistical noise was switched on.

The catch: that protective noise makes training three times slower and shaves a little accuracy. Coordinating five hospital IT departments is its own adventure. And no regulator has formally blessed this pattern yet, which keeps cautious institutions waiting.

Why you should care: the best medical AI needs experience no single hospital has — rare diseases, diverse patients — and that data is locked away for good reasons. If models can learn without the data ever moving, you eventually get the doctor-AI trained on millions of cases, and your records never leave the building.