Software & AI
Federated Learning
In one line
Training paradigm where model updates are computed on local data and aggregated centrally - raw data never leaves the site.
Definition
Federated learning enables multi-institution model training without centralizing PHI. Aggregated weights or gradients are shared, often with differential-privacy or secure aggregation safeguards.
Each site trains on its local data and shares only model updates, which a central server combines into a global model. This can support data diversity across hospitals while keeping data under local control, but it does not remove privacy risk, because model updates can sometimes leak information. For regulated devices, manufacturers still need to document data provenance, quality and representativeness for every site, which is harder when the data is not directly visible.
Why it matters
Increasingly common for medical imaging and EHR-based models; introduces new validation, governance, and IP questions.
Common pitfalls
- •Misinterpreting "decentralized" as inherently "private" or "secure" without additional safeguards.
- •Failing to implement robust data governance and access control mechanisms within each participating institution.
- •Underestimating the complexity of securing aggregation servers and communication channels.
- •Neglecting to address potential intellectual property rights associated with model contributions from various participants.
- •Assuming that federated learning automatically satisfies regulatory requirements for data anonymization or de-identification.
Sources
- NIH MIDRC · NIH
- MDCG Software Guidance · MDCG
- FDA - Software as a Medical Device (SaMD) · FDA