Hospitals want their AIs to learn from each other without sharing patient records. A new preprint proposes how.

Hospitals want better clinical AI for tasks like flagging patients at risk of readmission or sepsis, but they cannot legally or ethically ship patient records to each other. A new preprint proposes swapping what travels between hospitals during collaborative training: not the model updates, but the lessons a local AI agent learns while working on its own data.

The framework, called FedEHR-Agents, has two parts. Each hospital runs its own autonomous agent that preprocesses its electronic health record data and builds models locally, refining its work through past-task memory, task-specific evaluation, and prompt refinement. A central server then aggregates the lessons, what the authors call "clinical modeling experience," from heterogeneous hospitals and distills them into shared meta-prompts that each local agent can pull back in.

That is a different unit of collaboration than conventional federated learning, which shares model parameters or gradients. Here, the shared object is the distilled experience and prompt guidance. The authors report consistent gains over local and federated baselines on multi-hospital EHR benchmarks, and they report robustness across different federation scales and LLM backbones.

The ceiling is real. FedEHR-Agents is an arXiv preprint, not a peer-reviewed paper, with no patient outcomes, no deployment, and no independent replication. The paper does not enumerate which hospitals or datasets were used, only that the benchmark spans "diverse clinical prediction tasks." Whether experience-centric federation is meaningfully different from parameter-centric FL in practice is the question the field now has to answer.