Hospitals can now train AI models together without sharing patient data — if they speak the same medical language

Researchers built a system that lets hospitals in different countries train shared AI models on their patient data without sending that data anywhere. The trick: the system learns to translate between different hospitals' medical coding systems automatically, so a diagnosis coded one way in Boston matches the same diagnosis coded differently in Paris.

Right now, hospitals can't easily pool their data for research because each one codes diseases, treatments, and outcomes in its own way — and sharing patient records across institutions is legally and ethically fraught. This system solves the translation problem without requiring hospitals to adopt a single standard or send raw data outside their walls. That matters because clinical AI models trained on more diverse data tend to work better on real patients, and hospitals have been stuck choosing between privacy and scale. The catch: this is a proof-of-concept tested on seven institutions. Whether hospitals actually adopt it depends on whether it works as well in messy real-world deployments as it does in research settings, and whether the privacy gains hold up under regulatory scrutiny.