The best training data for medical artificial intelligence (AI) is usually the data that cannot leave the hospital. Patient scans, waveforms, and records sit inside hospital systems, wrapped in privacy law and legitimate caution, and that is exactly where the strongest signal lives. Federated learning in healthcare exists to resolve that contradiction. It lets teams train a shared model across many hospitals without any patient data leaving any building.
This is the short version of Avania’s white paper on federated learning for medtech AI. It covers what the method is, why it matters now, the privacy reality most vendors skip, and, most of all, how a model trained this way supports a U.S. Food and Drug Administration (FDA) Predetermined Change Control Plan (PCCP), the mechanism that lets a cleared model keep learning without a new submission for every update.
What federated learning is, in plain terms
The name sounds complicated. The mechanics are not. Instead of collecting patient data in one place to train a model, federated learning sends the model to the data.
A coordinating server holds a starting version of the model and sends a copy to each participating hospital. Each site trains that copy on its own local data, behind its own firewall. Training produces model updates, small adjustments to the model’s internal settings, not copies of the records. Each site returns only those updates, and the server averages them into an improved shared model and sends it back out. The loop repeats until the model converges.
- The model travels to each hospital. The patient data never leaves.
- Only model updates, not patient records, are sent back.
- A coordinating server averages the updates into one shared model.
The result is a model that learned from populations, scanners, and practice patterns it never saw. What traveled was math about the model, not the medical records.
Why federated learning in healthcare matters now
Two facts explain the timing. The first is where medical AI is being built. The FDA now lists more than 1,000 AI-enabled medical devices authorized for marketing, and roughly three-quarters of them sit in radiology (FDA, 2025). That concentration is a clue: models get built where a single institution can assemble enough labeled data in one place. The harder and more valuable problems that span many sites, rarer conditions, and diverse populations stall, because the data needed to train a generalizable model is scattered across organizations that cannot easily pool it.
The second fact is that federation works. In the EXAM study, published in Nature Medicine, a federated model that predicts how much oxygen a patient with COVID-19 will need was trained across 20 institutions with no patient data leaving any hospital. It improved average performance by roughly 16 percent and, more to the point, improved generalizability by roughly 38 percent compared with models trained at a single site (Dayan et al., 2021). An earlier multi-institution brain-tumor study reached about 99 percent of the accuracy of a model trained on centrally pooled data, again without any hospital sharing its patients (Sheller et al., 2020).
Bottom line: For the right problem, federated learning can approach the accuracy of pooling the data while sidestepping the exact reason the data could not be pooled.
The privacy reality most vendors skip
Federated learning improves the privacy posture of a multi-site program. It does not make privacy law or security disappear, and any vendor who says otherwise should worry you. Two points deserve candor.
- The legal frame stays. Keeping the raw records at each site reduces the data transfers regulators scrutinize most. It does not remove the need for a legal basis, institutional review where applicable, and data use agreements between the coordinator and each site. What changes is the shape of those agreements, not their existence.
- Model updates are not automatically safe. Under certain conditions, an adversary can infer information about the underlying training data from the updates alone, through model inversion or gradient-leakage attacks (Kaissis et al., 2020).
This is why serious deployments do not stop at “the data stayed home.” They add secure aggregation, so the coordinator sees only the combined update and never any single site’s contribution, and differential privacy, which adds calibrated noise to limit what any update can reveal. The governance question is not whether to use these protections but which combination the threat model and the regulators require.
The regulatory bar does not lower, it moves
A model trained across sites clears the same bar as any AI-enabled device, and in one respect a higher one. Federated training changes how the evidence is generated. It does not lower the standard of evidence. The FDA’s Good Machine Learning Practice (GMLP) guiding principles ask that training data represent the intended population and that performance be characterized across clinically relevant subgroups (FDA, 2021). Federation is a genuine strength here, because it draws on more sites and more diverse populations than any single hospital can. That strength only counts if it is demonstrated on a locked, independent test set that was never used in training. Prove generalizability. Do not assume it.
Federated learning and the FDA PCCP
The requirement that decides whether a distributed model is worth building is change control. Federated systems make continued, multi-site training easy to imagine, which is exactly why the plan for how the model may change has to be settled up front. The FDA’s Predetermined Change Control Plan (PCCP) guidance, finalized in December 2024, lets a manufacturer pre-specify in the original submission which future updates are pre-authorized and how they will be validated, so anticipated changes do not each trigger a new submission (FDA, 2024). A PCCP has two working parts: pre-specifications that describe which modifications are anticipated, and an algorithm change protocol that fixes the method, the data, and the metrics for making and verifying each change.
Federated learning is a natural mechanism for that update process. Retraining the model across a defined network of hospital sites, while the patient data stays local, gives the disciplined, repeatable, and auditable pipeline that an algorithm change protocol is meant to describe. It also produces the evidence the FDA is looking for. Because the model is trained and validated across many institutions rather than one, generalizability is built into how it is made, and because the network can refresh the model on new multi-site data and watch performance site by site, the postmarket monitoring a PCCP depends on has real signal from the places the model is actually used.
The discipline still governs, and that is the point. Every update has to stay within the pre-specified modifications and the cleared intended use, and every update has to be validated on an independent, locked test set with methods that give the FDA equivalent assurance of safety and effectiveness. Where a proposed change would alter the validation method itself or reach beyond the anticipated modifications, that is a new regulatory question, not a routine update, and the plan should say so. The move that saves time later is to write the pre-specifications and the algorithm change protocol so that the federated update process and the regulatory limits are a single plan, not two that have to be reconciled after the fact.
How Avania builds and defends distributed AI
Avania’s Digital and AI Center of Excellence brings AI and machine learning engineering and FDA and European Union regulatory strategy under one roof, which is the combination distributed AI actually requires. Most firms do one or the other. For teams weighing a federated or otherwise privacy-preserving approach, we help decide whether federation is the right tool, design the multi-site evidence and governance so the data can stay put, and shape the GMLP and PCCP strategy a distributed model will be held to.
We work hands-on. Standing up the full federated loop in code, not slides, is what surfaces the questions a leadership team has to answer before funding an enterprise program: where secure aggregation sits in the pipeline, who holds the locked test set, and how the algorithm change protocol maps onto the retraining you actually intend to run. Building the pipeline and defending it in front of a regulator are the same job, and we do both. Avania supports and helps position the pathway; regulators clear the product.
The bottom line
Federated learning in healthcare is not a way around privacy law or the validation bar. It is a way to build medical AI on the data that matters most, the data that cannot leave the hospital, while keeping it where it belongs. Ask three questions before committing. Is your best model genuinely blocked because the data cannot be centralized? Does your privacy design go beyond “the data stayed home” to include secure aggregation and, where warranted, differential privacy? And have you planned the regulatory path from the start, with a locked test set and a Predetermined Change Control Plan? The full white paper works through each one.
Reference
1 Dayan, I., Roth, H. R., Zhong, A., Harouni, A., Gentili, A., Abidin, A. Z., … Li, Q. (2021). Federated learning for predicting clinical outcomes in patients with COVID-19. Nature Medicine, 27(10), 1735-1743. https://doi.org/10.1038/s41591-021-01506-3
2 Sheller, M. J., Edwards, B., Reina, G. A., Martin, J., Pati, S., Kotrotsou, A., … Bakas, S. (2020). Federated learning in medicine: Facilitating multi-institutional collaborations without sharing patient data. Scientific Reports, 10, 12598. https://doi.org/10.1038/s41598-020-69250-1
3 Kaissis, G. A., Makowski, M. R., Ruckert, D., & Braren, R. F. (2020). Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6), 305-311. https://doi.org/10.1038/s42256-020-0186-1
4 U.S. Food and Drug Administration. (2021). Good machine learning practice for medical device development: Guiding principles. https://www.fda.gov/medical-devices/software-medical-device-samd/good-machine-learning-practice-medical-device-development-guiding-principles
5 U.S. Food and Drug Administration. (2024). Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions. https://www.federalregister.gov/documents/2024/12/04/2024-28361/marketing-submission-recommendations-for-a-predetermined-change-control-plan-for-artificial
6 U.S. Food and Drug Administration. (2025). Artificial intelligence-enabled medical devices [Device list]. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices