Federated learning is a machine-learning method that trains one shared model across many institutions without moving their data. Each site trains locally on its own patients; only model updates — never records — travel to a coordinator, which combines them into a single consensus model. As of September 2026.
The term was coined in 2017 to describe a decentralized alternative to collecting everything in one place: leave the training data where it lives, learn locally, and aggregate the locally-computed updates into a shared model 1. In healthcare the same idea answers a specific problem — the most valuable training data sits in separate hospital systems that cannot legally or practically merge their records.
Why it matters in healthcare
Patient data is both the fuel for clinical AI and the thing that regulation, consent, and institutional trust most tightly restrict. A model trained on one hospital's population tends to generalize poorly to another's. Federated learning lets institutions gain insights collaboratively — in the form of a consensus model — without moving patient data beyond the firewalls of the sites where it resides 2. That reframes the classic trade-off: you no longer have to choose between a bigger, more representative training set and keeping data under local governance.
How it works in practice
A coordinator sends the current model to each participating site. Every site trains on its own data for a round, then returns only the updated model parameters. The coordinator averages those updates into a new shared model and repeats. Raw records never leave the building. Because sites differ in scanner hardware, coding habits, and patient mix, real deployments add steps to handle that heterogeneity, and they layer on privacy protections described below.
Where it appears today
The evidence base is now built on real multi-site systems, not demonstrations. A multi-institutional study showed federated models reaching up to 99% of the accuracy of a model trained on the same data pooled centrally — without any site sharing its data 3. During the pandemic, the EXAM model was trained across 20 institutes worldwide to predict the oxygen needs of patients with COVID-19, achieving an average AUC above 0.92 and a 38% average gain in generalizability over models trained at a single site 4. The largest reported effort assembled data from 71 sites across six continents — 6,314 glioblastoma cases — to train a rare-cancer boundary detector that no single institution had the volume to build alone 5.
Common misunderstandings
The most consequential error is treating federated learning as automatic privacy. Keeping raw records local removes the biggest exposure, but the model updates that do travel can still leak information about the training data. Serious systems therefore pair the federated setup with secure aggregation, so the coordinator sees only the combined update, and differential privacy, which adds calibrated noise to bound what any update reveals 6. A second misunderstanding is assuming a federated model equals a centralized one; site heterogeneity, uneven data quality, and coordination overhead mean results have to be validated, not assumed. Federated learning widens the training set — it does not retire the need for external validation.
Related terms
Federated learning sits alongside other techniques for training on sensitive records. Synthetic patient data generates artificial records that stand in for real ones, while de-identification versus anonymization governs when stripped data is still legally personal. Many federated projects now train or fine-tune a foundation model rather than a single-task network. For the broader numbers on where clinical AI has spread, see our AI in healthcare statistics.
We revisit this page on a yearly cycle, and sooner if a new multi-institutional deployment, a regulator's guidance on federated training, or a demonstrated attack on shared updates changes the picture.