The Data Paradox in Multi-Site Clinical Research
Multi-site clinical studies face a structural tension. Statistical power and generalizable findings require large, diverse patient cohorts, but the patient-level data needed to build that cohort is legally and ethically bound to the institution that collected it. The GDPR restricts the transfer and pooling of health data, a special category of personal data, unless a valid legal basis and adequate safeguards are in place. Switzerland's revised Federal Act on Data Protection (nDSG) imposes parallel obligations, treating health information as sensitive personal data and layering on physician confidentiality duties under the Swiss Criminal Code, alongside sector rules such as the Health Insurance Act and the Electronic Patient Record Act. Guidance from Swiss data protection specialists notes that healthcare organizations are increasingly expected to keep data under Swiss residency or, at minimum, transparent EU/EEA hosting with documented sub-processors.
A hospital in Basel, a university clinic in Lyon, and a research center in Boston may each hold pieces of the same rare disease puzzle, yet no single institution can legally or ethically ship its patient records to a central server for pooled analysis. Traditional centralized machine learning, which assumes all training data lives in one place, runs directly into this constraint. The historical result has been smaller, single-site models with limited statistical power and a persistent risk of bias toward whichever population happened to be available.
How Federated Learning Works
Federated learning inverts the usual data flow: instead of moving patient data to a central location, the algorithm travels to the data. A coordinating server distributes a shared model architecture to each participating site. Each site trains that model locally, using only its own patient records, which never leave its firewall. Only the resulting model parameters, weights and gradients, are sent back to the coordinator, which aggregates the updates from all sites into an improved consensus model. That consensus model is then redistributed for another round of local training, repeating until performance converges.
- Raw patient-level data stays inside each institution's own infrastructure and governance boundary.
- Only model updates, not source records, cross institutional or national borders.
- Additional safeguards such as differential privacy and secure aggregation can be layered on to reduce the residual risk that model updates could be reverse-engineered to reveal information about individual patients.
As researchers writing in npj Digital Medicine put it, federated learning lets institutions "train algorithms collaboratively without exchanging the data itself," because participants "never directly access data from other institutions and only receive model parameters that are aggregated over several participants." The same authors caution this is not an automatic privacy guarantee: gradient inversion attacks can, in principle, extract information from shared updates, which is why serious deployments pair federation with technical controls and clear governance agreements rather than treating the architecture alone as sufficient compliance.
Evidence from Real Deployments
The approach has moved well past proof of concept. The Federated Tumor Segmentation (FeTS) initiative, published in Nature Communications, coordinated 71 sites across six continents to train a glioblastoma boundary-detection model on 6,314 patient cases without centralizing any imaging data, the largest federated learning study of its kind to date. The resulting consensus model showed substantial gains over the initial public baseline model, including improvements of 27 percent, 33 percent, and 16 percent on enhancing tumor, tumor core, and whole tumor sub-compartments respectively on local validation data, and it performed as well as or better than models trained on data from the four largest individual contributing institutions.
In oncology research, Owkin used its Substra federated learning framework to train deep learning models on digital histopathology slides from four French hospitals (Institut Curie, Centre Leon Berard, Gustave Roussy, and IUCT Oncopole), covering 650 triple-negative breast cancer patients, one of the largest cohorts assembled for this kind of analysis, to predict response to neoadjuvant chemotherapy without any hospital's pathology images leaving its own systems, with results published in Nature Medicine.
Pharmaceutical R&D offers a parallel case outside direct patient care. The MELLODDY consortium brought ten pharmaceutical companies (including Novartis, Bayer, Boehringer Ingelheim, and AstraZeneca) together to federate more than 2.6 billion confidential activity data points across 21 million compounds, without any company exposing its proprietary chemical library to competitors. Published results in the Journal of Chemical Information and Modeling reported that all participating partners saw positive effects on model quality, with median AUC-PR improvements exceeding 2.5 percent for classification models and the strongest decile of models gaining at least 20 percent, evidence that federation can unlock statistical power even among commercial rivals with strong incentives to keep data siloed.
Models trained by federated learning can achieve performance comparable to models trained on centrally hosted data sets, and superior to models that only see isolated single-institution data.
NVIDIA's Clara federated learning platform has extended this pattern to mammography and other imaging tasks, coordinating multiple hospital partners to jointly improve models while each site's imaging data stays behind its own firewall, an architecture increasingly cited as a template for privacy-preserving collaboration in radiology.
Where a Firm Like ANG Associates Fits
Federated learning does not remove the need for careful governance, it relocates it. Every participating site still needs a validated local environment, a documented data processing agreement, a clear model of which entity is controller and which is processor under GDPR or nDSG, and evidence that the aggregation server and update protocol have been assessed for residual re-identification risk. In GxP-regulated environments, the federated model itself, its training pipeline, and its validation evidence must satisfy the same computerized system validation expectations as any other tool influencing clinical or regulatory decisions. This is where a Basel-based consulting firm with both AI strategy and GxP validation expertise, like ANG Associates, can help: translating a promising federated architecture into a data governance framework, a validation and change-control plan, and a delivery roadmap, often run through SAFe or Agile ceremonies across multiple sponsor and site teams, that a Swiss or EU life sciences organization can actually defend to an auditor, an ethics committee, or a data protection authority.
Sources
- Pati, S., Baid, U., Edwards, B. et al. "Federated learning enables big data for rare cancer boundary detection." Nature Communications, 2022.
- "Nature Medicine publishes breakthrough Owkin research on the first ever use of federated learning to train deep learning models on multiple hospitals' histopathology data." EurekAlert! (Owkin press release), 2023.
- Guinney, J. et al. (MELLODDY Consortium). "MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising Proprietary Information." Journal of Chemical Information and Modeling, 2024. https://pubs.acs.org/doi/10.1021/acs.jcim.3c00799 (open-access copy:
- Rieke, N., Hancox, J., Li, W. et al. "The future of digital health with federated learning." npj Digital Medicine, 2020.
- "Federated Learning powered by NVIDIA Clara." NVIDIA Technical Blog.
- "Data Protection in Swiss Medical Practices, nDSG, Patient Records, EPR." SIDD Insights.
- "A scoping review of the governance of federated learning in healthcare." npj Digital Medicine, 2025.