Federated Learning in Healthcare

Federated learning allows multiple organizations to train a shared model without pooling their raw data in one central repository. Sites send model updates or gradients to an aggregation process, then receive an updated model. A Nature Medicine study of the EXAM model showed a multi-institutional workflow for predicting clinical outcomes in patients with COVID-19 without exchanging underlying datasets. That study demonstrates one implementation, not a universal guarantee of privacy or performance. Data remain distributed, but information can still leak through model updates, metadata, or poorly secured infrastructure. Federated learning does not automatically solve differences in coding, measurement, patient populations, or missing data across hospitals. Some sites may dominate updates if their datasets are larger, and models can perform poorly for underrepresented populations. Secure aggregation, privacy-preserving techniques, access control, and governance may be needed. Before collaboration, define the task, local data standards, update protocol, security controls, and responsibilities for monitoring. Evaluate the final model at sites that did not contribute to training and report performance by institution and subgroup. Confirm that participants and institutions have appropriate governance for data use. Federated learning can make collaboration possible, but it does not replace privacy assessment, external validation, or clinical review. Explain whether model updates are aggregated centrally, how participants can withdraw, and how security incidents will be handled. Sites should agree on the meaning of labels and on the escalation path when an institution’s data differ materially from others.