The patients AI misses
AI disease detection is only as good as the data it learns from, and that data is deeply uneven. Models trained at elite research centers inherit the demographics of those centers' patients, and 1 in 3 AI-assisted diagnoses has been found vulnerable to demographic bias (Nature Medicine, 2021). The populations left out are usually the ones already underserved.
The gap is geographic too. In the United States, 68.5% of counties, home to more than 50 million Americans, have no radiation oncology practice site (ASTRO, 2026). Rural patients travel farthest for care, and cancer mortality in rural areas now exceeds urban rates (American Cancer Society). Where detection AI could help most, the models have learned the least about those patients.
The usual answer, pooling patient records into a central database, is blocked by design. HIPAA, GDPR and CCPA restrict how protected health information moves; IRBs hold data use to its original consent scope; and no hospital wants to hand its records to a central warehouse that becomes a single point of failure for every participant's patients. So each hospital trains alone, on its own patients, and its model inherits its own population's blind spots.
The solution: Oncology HealthFL
Oncology HealthFL is Goose FL's application of the healthcare use case to cancer detection. The model travels to each hospital and trains locally behind its firewall. Hospitals return protected mathematical updates, not patient records; the updates are combined into a stronger shared model; and each hospital receives that model personalized to its own population, not a one-size-fits-all average.
Why personalization, not just federation
Standard federated learning (FedAvg) averages every hospital into one global model. A region's 20% mortality rate gets blended with another's 6%. The average fits nobody, and the communities already missing from the data are erased first. Personalized federated learning (pFedMe) lets every hospital keep a model tuned to its own patients while still learning from the whole network.
Every hospital receives the same averaged model. Structural differences between populations are treated as noise, and averaged away.
A model that learns one region's tumor-and-treatment pattern misjudges patients whose profiles differ.
Each hospital gets its own model, tuned to its own patients, while still benefiting from the network's shared learning.
The model understands the local community, not an average of six.
The evidence: 1,096 real patients, six very different populations
The proof of concept runs on TCGA-BRCA: 1,096 real breast cancer patients across 6 geographic regions. Same dataset, radically different clinical profiles, which is precisely the non-uniform reality federated learning has to handle:
| Region | Northeast | South | West | Midwest | Europe | Canada |
|---|---|---|---|---|---|---|
| Avg age | 58 | 58 | 59 | 56 | 64 | 55 |
| Mortality | 19% | 20% | 11% | 12% | 6% | 6% |
| Small tumors (T1c) | 31% | 21% | 13% | 13% | 22% | 2% |
| Large tumors (T3) | 8% | 16% | 17% | 14% | 10% | 14% |
| Lymph nodes (N1+) | 37% | 28% | 38% | 40% | 40% | 59% |
Clinical profiles vary profoundly by region, the non-IID reality that averaging destroys and personalization preserves.
On this data, with the full privacy stack active, the personalized model beats the global average, across 10 random seeds:
A proof of concept, not a clinical product: binary survival prediction from 39 clinical features, validating that personalized federated learning holds its advantage under privacy constraints on real, heterogeneous clinical data. The architecture is target-agnostic, the same stack can address treatment response, recurrence risk and screening risk with different labeled data.
See it run
The recorded walkthrough below shows Oncology HealthFL's federated training loop end to end, regions training locally, updates combining under the privacy stack, and the personalized results compared against the averaged baseline.
Privacy that holds at every layer
The architectural stack, every layer exercised in the proof of concept:
- pFedMe personalization, each hospital's model is tuned to its population while learning from the group (Moreau-envelope regularization; the server blends the global model with the average of personalized models each round).
- Secure multiparty computation, updates are combined in an encrypted state. No participant, and not even the operator, can inspect another hospital's contribution.
- Differential privacy, calibrated noise protects every patient's record from being reverse-engineered from model updates. Crucially, the noise is decoupled per center: standard DP applies the same noise everywhere, which quietly degrades the model for the smaller, rarer-population centers that need detection most. Goose FL's privacy-equity architecture adapts the noise to the signal.
- FHIR R4 / OMOP edge nodes, the model draws clinical data through the hospital's existing Epic or Cerner FHIR API and standardizes it via OMOP, entirely on hospital infrastructure.
Why the privacy design is an equity design
Standard differential privacy applies uniform noise to every participant. But minority and rural populations have smaller cohorts and weaker signal, a center with 5.6% cancer prevalence suffers far worse noise-to-signal than one with 19.9%. Uniform noise systematically degrades the model for the communities that need it most. Decoupled privacy budgets correct this: protection stays absolute, but the noise respects each population's reality.
The federated pattern's clinical proof point is established in medicine: a tumor-detection model trained across 71 hospitals improved detection by a third with zero patient records shared (Sheller et al., 2020), the evidence base this architecture builds on.
Learn more
- Oncology HealthFL, the full presentation
- Federated Learning 101, how federated learning works, from the privacy paradox to the shared model
- The Goose FL trust stack, the open core, trust layers and governance underneath this use case
- The Finance use case, the same infrastructure applied to AML/CFT detection