Federated Learning Boosts Cancer Survival Models Across Hospitals Without Sharing Patient Data
Three AI models tested on breast cancer data from multiple hospitals show federated learning outperforms local training
A new study from Moreno-Blasco et al. systematically evaluates federated survival analysis for breast cancer prognosis using real-world cross-institutional data. Three representative survival models—the classic Cox Proportional Hazards, neural-network-based DeepSurv, and ensemble-based Random Survival Forest (RSF)—were compared across centralized, local, and federated training setups. The federated setup simulated naturally heterogeneous hospital clients with varying patient distributions. Three optimization strategies (FedAvg, FedProx, FedAdam) were tested for gradient-based models.
The results show that federated learning (FL) consistently beats local training, and in some cases even matches or exceeds centralized performance—a crucial finding since centralizing patient data is often illegal under privacy regulations like HIPAA and GDPR. RSF emerged as the top performer overall, delivering the best balance of discrimination (how well it separates patients who survive vs. those who don't), calibration (accuracy of predicted survival probabilities), and robustness across diverse client distributions. Among FL optimizers, FedAvg and FedProx were significantly more stable and effective than FedAdam. The paper also derives decision-oriented guidelines to help healthcare AI teams select the right model and training paradigm based on their specific data diversity, privacy requirements, interpretability needs, and computational resources. This work brings practical clarity to deploying privacy-preserving survival models in clinical settings.
- Random Survival Forest (RSF) outperformed Cox and DeepSurv models under federated learning on heterogeneous breast cancer data
- Federated learning matched or exceeded centralized training performance while respecting patient privacy regulations
- FedAvg and FedProx optimization strategies proved more stable and effective than FedAdam for gradient-based survival models
Why It Matters
Enables hospitals to collaboratively train accurate cancer survival models without sharing sensitive patient data, improving prognostics across institutions.