Mapping Federated Learning Use Cases Across Clinical Medicine

The most useful way to judge federated learning healthcare AI use cases is not to begin with the architecture diagram. It is to ask where a clinical consortium actually trained a model without first pooling patient data, how many sites participated, what the model was asked to do, and whether the result looked clinically credible outside one institution.

The EXAM study is still the cleanest entry point. During the COVID-19 pandemic, 20 institutions across four continents used federated learning to train a model for predicting future oxygen requirements in symptomatic patients. The model reached AUC values above 0.92, improved performance by 16% compared with local models, and showed a 38% generalizability gain over models trained at a single site.[1]

That is the kind of claim that matters operationally. Twenty hospitals did not merely agree in principle that data sharing was hard. They participated in a workflow where local data stayed local, model updates moved, and a clinically recognizable deterioration task produced better cross-site performance than isolated local training. It does not settle every privacy, liability, or deployment question. It does show why health systems keep returning to federated learning when conventional multi-site pooling is legally slow, politically fragile, or technically unrealistic.

For readers who need the mechanics first, the federated learning glossary entry covers the basic model-update pattern and privacy mechanisms. This article stays with the clinical map: where federated learning has been tested, what evidence is strongest, and where the practical gaps remain.

The Evidence Base Is Broader Than the Headline Cases, but Still Early

A 2024 systematic review by Teo and colleagues is the best load-bearing source for the field because it prevents a few successful deployments from standing in for the whole evidence base. The review examined 612 articles on federated learning in healthcare. Medical imaging accounted for 41.7% of studies, electronic health record data for 23.7%, and genomics for only 2.3%.[2]

Across studies comparing federated models with centralized training, federated learning generally achieved 95–98% of centralized performance.[2] That is a meaningful translational result because centralized training is often the method that multi-site healthcare research cannot easily use. A model that approaches centralized performance while avoiding raw-data pooling can make a study possible rather than merely more elegant.

The same review also keeps the enthusiasm in bounds. Only 5.2% of included articles described real-world deployment, and the review’s evidence window ran through August 2023.[2] By Q3 2026, that figure is better treated as a floor estimate from the published literature rather than a live census of deployed systems. Still, it is a sobering floor: most federated learning work in healthcare remained experimental, simulated, or benchmark-oriented in that evidence window.

Radiology Shows Why Federated Learning Became Clinically Plausible

Radiology is where federated learning has had the easiest path to credibility, partly because imaging already has mature AI benchmarks and partly because image data are large, identifiable in practice, and often difficult to move between institutions. The Teo review’s 41.7% imaging share is not a minor skew; it is the dominant shape of the field.[2]

EXAM sits inside that imaging-heavy pattern but is more clinically interesting than a standard classification benchmark. It combined chest imaging with the urgent pandemic task of predicting clinical deterioration, and it did so across sites with different patient populations and operational settings. The reported AUC above 0.92 would be less persuasive if it came from one institution; the 20-institution, four-continent design is why the result still gets cited.[1]

Brain tumor segmentation is another natural fit. Tumor imaging differs by scanner, protocol, annotation style, and referral pattern. A single academic center may have deep expertise but not enough diversity to represent the range of disease appearance and acquisition conditions a model will face elsewhere. Columbia’s 2025 workshop highlighted multi-institutional federated learning work in brain tumor segmentation, along with other clinical use cases, as evidence of how institutions are trying to collaborate without sending raw clinical data into one shared repository.[3]

The same heterogeneity that makes federated learning attractive also makes it technically and clinically difficult. Non-IID data are not a footnote in imaging; they are the daily condition of radiology practice. A hospital with one scanner vendor, one reconstruction protocol, and one referral base will not look like another hospital across town, much less one on another continent. Federated learning can expose a model to that variation without centralizing images, but it does not automatically make labels consistent, scanner artifacts irrelevant, or evaluation fair across sites.

That distinction matters for adoption. A radiology department asked to join a federated project still has to map local data, confirm labels, run local infrastructure, monitor updates, and decide whether the final model is safe for its own patients. The model may never have moved the images, but the burden of local data quality remains very much on the contributing institution.

Oncology: Faster Collaboration, Slower Data Harmonization

Oncology has a different kind of urgency. The data needed for recurrence prediction, treatment response, or tumor characterization may be distributed across hospitals, pathology systems, imaging archives, oncology registries, and follow-up records. Pooling those data can take longer than the model development itself.

The Kakao Healthcare and Google Cloud breast cancer recurrence project makes that operational point unusually concrete. Sixteen Korean hospitals participated, the project timeline was reduced from an expected two years to four months, and the federated model achieved an AUC of 0.8482.[4]

The attractive headline is the timeline reduction. The more transferable lesson is that the hardest work was not pressing a federated training button. Kakao and Google describe harmonizing data across 16 hospital systems as the most difficult part of the project.[4] Anyone who has tried to reconcile recurrence definitions, surgery dates, pathology fields, medication records, and follow-up intervals across oncology sites will recognize the shape of that problem.

Federated learning helped reduce the need to negotiate raw-data transfer, but it did not remove the need to agree on what the outcome meant. A recurrence model cannot be clinically interpretable if one hospital codes recurrence after imaging suspicion, another after biopsy confirmation, and another after systemic therapy restart. The method protects where data sit; it does not repair inconsistent clinical meaning.

Brain tumor segmentation also belongs in the oncology map, although it overlaps with radiology. Its value is precisely that overlap: oncology needs longitudinal disease characterization, radiology supplies much of the measurable phenotype, and neurosurgery or radiation oncology may judge whether a segmentation is useful. Federated learning can support this kind of cross-institutional model development, but the downstream clinical governance is broader than one department.

Drug Discovery Solves a Different Collaboration Problem

Drug discovery is often grouped with healthcare AI, but the federated learning problem is different. Hospitals are usually trying to avoid exposing patient-level clinical data. Pharmaceutical companies are trying to collaborate without exposing proprietary compound libraries and assay data.

MELLODDY is the flagship case. The consortium brought together 10 pharmaceutical companies and used more than 2.6 billion data points covering more than 21 million compounds to train models without compromising the companies’ proprietary chemical libraries.[5]

That is not a hospital deployment, and it should not be evaluated as though it were one. The clinical question is not whether a bedside team can act on the output tomorrow. The collaboration question is whether competing organizations can improve predictive modeling while preserving commercially sensitive data boundaries. MELLODDY demonstrated that this form of cross-pharma federated learning can be made to work at a scale that would be hard to reproduce through conventional data sharing.[5]

For readers comparing this area with AI drug discovery companies more generally, the distinction is important: federated learning is not a discovery strategy by itself. It is a collaboration architecture that may allow richer model training when no single company is willing to contribute its compound library to a pooled dataset. The broader AI drug discovery company landscape includes many models and business approaches that do not depend on FL.

Ophthalmology, Neurology, ICU, and Cardiology Extend the Map

The evidence outside radiology and oncology is thinner, but it is not absent. Columbia’s workshop described federated learning applications in retinopathy of prematurity, traumatic brain injury deterioration prediction across U.S. and U.K. sites, and cardiac MRI, among other use cases.[3] These are not interchangeable examples. Each uses a different clinical workflow, data type, and adoption pathway.

Ophthalmology is a logical setting because screening problems such as diabetic retinopathy and retinopathy of prematurity combine image interpretation with access constraints. A model that generalizes across neonatal units or screening programs could matter, especially where specialist availability is uneven. The current public evidence, however, is not as deep as the imaging literature overall, and the cited sources do not support treating ophthalmology as equally mature.

Neurology and ICU use cases move federated learning away from static images and toward acute deterioration prediction. Traumatic brain injury is a good example because deterioration risk depends on time-sensitive clinical variables, imaging findings, interventions, and site-level care patterns. A federated approach may help capture that variation across hospitals, but it also increases the burden of temporal alignment and outcome definition. In ICU data, the timestamp can be as important as the value.

Cardiology sits somewhere between the imaging-heavy and EHR-heavy sides of the field. Cardiac MRI use cases resemble radiology in their dependence on image acquisition and segmentation quality. Predictive cardiology models across hospital systems raise the more familiar EHR problems: coding variation, medication history completeness, device data availability, and follow-up capture. The specialty is active, but the evidence summarized here does not justify presenting it as a leading deployment area.

What the Specialty Map Leaves Out

The map is conspicuously sparse in genomics. Teo and colleagues found that genomics accounted for only 2.3% of healthcare federated learning studies.[2] That is surprising only if one looks at potential rather than implementation. Genomic data are sensitive, large, heterogeneous, and institutionally guarded, which makes FL attractive in theory. In practice, the published healthcare FL literature through the review window did not make genomics a major evidence cluster.

EHR-based work is more common, at 23.7% of studies, but it should not be mistaken for easy deployment.[2] Tabular clinical data carry their own version of the scanner problem: local coding practices, missingness patterns, care pathways, and documentation incentives. A federated model trained across hospitals may be exposed to more variation, but it still inherits the measurement habits of each participating system.

That matters for equity as well as performance. Sites differ demographically, geographically, and economically. Distributed training can broaden the data environment, but it can also hide subgroup failure if evaluation is reported only as an aggregate. Teams working in clinically diverse networks should treat site-stratified and subgroup-stratified evaluation as part of the model’s safety case, not an optional appendix. The same concern sits behind broader work on algorithmic bias in clinical AI.

Privacy Is Not Automatic Just Because Training Is Federated

The phrase “we never moved the data” is useful, but it is not a complete privacy argument. In the Teo review, only 27.7% of studies used formal privacy-enhancing technologies such as differential privacy, homomorphic encryption, or secure multiparty computation.[2] That means many healthcare FL studies relied on the distributed training setup itself rather than layering stronger privacy protections on top.

That distinction should be explicit in project proposals. Federated learning can reduce raw-data transfer, which is often the largest barrier to multi-site collaboration. It does not, by itself, answer every question about model update leakage, membership inference, audit rights, or how much information a trained model may reveal about a contributing site. The privacy claim has to match the implementation.

The governance gap is just as important. Among the 32 real-world deployment studies identified in the systematic review, none documented IP-rights distribution between sites.[2] That omission is easy to overlook during a pilot and hard to repair after a model performs well. If a hospital contributes data, local infrastructure, clinical expertise, and patient population diversity, it will reasonably ask who owns the resulting model, who can commercialize it, who maintains it, and who is accountable when it changes.

This is where the translational gap in healthcare AI usually becomes visible. The model can be technically impressive and still stall because contracting, monitoring, reimbursement, liability, and workflow integration were treated as afterthoughts. Federated learning removes one major obstacle to collaboration; it does not remove the rest of the deployment pathway. That broader problem is familiar across AI systems that struggle to move from research into clinical use.

A Practical Readiness Test for Federated Learning Projects

The strongest federated learning healthcare AI use cases share a practical pattern: the collaboration barrier is real, the data cannot easily be pooled, the task benefits from multi-site variation, and the consortium is willing to do the unglamorous work before training starts. A project does not become clinically serious because it uses federated learning. It becomes serious when its design matches the specialty, data modality, and deployment pathway.

• Real deployment precedent: Has a similar FL task been tested outside simulation in this specialty or data modality?
• Site diversity: Are there enough participating institutions to test generalizability rather than reproduce one hospital’s habits?
• Data harmonization: Are labels, time windows, inclusion criteria, and local schemas agreed before training begins?
• Privacy implementation: Does the project use formal privacy-enhancing technologies, or is it relying only on local data retention?
• Ownership and accountability: Is there a written answer for IP rights, model maintenance, commercialization, monitoring, and withdrawal?
• Clinical evaluation: Will results be reported by site and relevant subgroup, not only as pooled performance?

By that standard, federated learning is no longer merely a privacy-preserving research concept. It has credible multi-site evidence across radiology, oncology, ophthalmology, neurology, cardiology, and drug discovery. It often approaches centralized-training performance, and in cases such as EXAM it has shown measurable generalizability gains over single-site models.[1][2]

It is also not mature enough to treat as default clinical infrastructure. The published evidence remains concentrated in imaging, real-world deployment was uncommon in the systematic review window, formal privacy mechanisms were used in only a minority of studies, and IP-sharing was undocumented across the real-world deployments reviewed.[2] The working map is therefore calibrated rather than promotional: federated learning is most convincing today where collaboration barriers are real, data cannot practically be pooled, and the consortium is prepared to do the coordination work before claiming clinical impact.

References

1. Federated learning for predicting clinical outcomes in patients with COVID-19, Nature Medicine, 2021
2. Federated learning in healthcare: A systematic review, Cell Reports Medicine, 2024
3. Federated Learning Columbia Workshop Showcases Future of AI Collaboration in Healthcare, Columbia University Vagelos College of Physicians and Surgeons, 2025
4. Kakao Healthcare Uses Federated Learning on Google Cloud to Accelerate Breast Cancer Recurrence Prediction, Google Cloud, 2024
5. MELLODDY: Cross-Pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising Proprietary Information, Journal of Chemical Information and Modeling, 2023