Federated Learning — Glossary

Federated learning trains a shared model across multiple participants (hospitals, research sites, institutions) without requiring any participant to share their raw data. Instead of centralizing data in one place, the model is distributed to each site, trained locally on that site’s data, and only the model updates (gradients or weights) — not the underlying data — are sent to a central aggregator. The central model incorporates all sites’ updates and distributes the improved model back.

Why it matters for medical research:

Clinical data is subject to patient privacy regulations (HIPAA in the US, GDPR in Europe) that often prohibit raw data from leaving an institution, or require complex data-sharing agreements that can take months to negotiate. Federated learning allows multi-site research studies to train on combined data from dozens of hospitals without any data transfer.

Applications in medicine include:

Training imaging AI models on radiology data from multiple hospital systems
Building disease prediction models from electronic health records across institutions
Rare disease research where no single site has sufficient cases

Limitations:

Gradient leakage attacks: Research has shown that model gradients can, under some conditions, be used to partially reconstruct training data. Federated learning reduces privacy risk but does not eliminate it — additional techniques like differential privacy are often paired with it.
Statistical heterogeneity: If different sites have different patient populations or imaging equipment, local models may diverge, making aggregation less effective.
Communication overhead: Many rounds of gradient communication are required, which can be slow over network-limited connections.
IRB and governance still apply: Federated learning does not remove the need for institutional review — it changes the data flow, not the ethical obligations.

Related guide: Health & Medicine