Federated Learning for Multi-Institution Oncology Collaboration

Context

Healthcare institutions often need to collaborate to build robust predictive models in oncology. While each hospital holds valuable patient data, strict privacy regulations such as GDPR prevent direct data sharing across sites.

Challenge

Centralizing medical data is not feasible due to legal, ethical, and governance constraints. As a result, institutions struggle to access sufficiently large and diverse datasets to train reliable models.

Solution with Fed-BioMed

Fed-BioMed enables a federated learning approach where each institution trains models locally on its own data. Only model parameters or updates are shared, ensuring that sensitive patient data never leaves the hospital.

Key aspects of the deployment:

• Local training on institutional data without data transfer
• Secure aggregation of model updates across sites
• Governance controls allowing institutions to validate and approve training workflows
• Reproducible experiments across distributed environments

Results

This approach allows institutions to collaboratively train high-quality models while maintaining full control over their data. Federated learning enables access to broader data diversity, improving model robustness without compromising privacy or compliance.

Typical Participants

• University hospitals
• Oncology centers
• Clinical research institutions