Federated Survival Analysis in Healthcare: A Multi-Model Evaluation on Cross-Institutional Heterogeneous Breast Cancer Data

Research Question and Approach

This study addresses the challenge of performing survival analysis—predicting time-to-event outcomes like disease progression or death—in healthcare environments where data are siloed across institutions due to privacy regulations. The authors evaluate whether federated learning (FL) can effectively train survival models without pooling raw patient records. They compare three distinct modeling approaches: the classical Cox Proportional Hazards (CoxPH) model, the deep-learning-based DeepSurv, and the ensemble-based Random Survival Forest (RSF). These models were tested across centralized, local, and federated training paradigms using the Fed-TCGA-BRCA breast cancer dataset, which features natural cross-institutional heterogeneity in patient demographics and survival profiles.

Federated Training Strategies

For the gradient-based models (CoxPH and DeepSurv), the researchers assessed three federated optimization strategies: FedAvg, FedProx, and FedAdam. FedAvg serves as the standard baseline, while FedProx introduces a proximal term to mitigate model divergence caused by heterogeneous data distributions across clients. FedAdam applies adaptive moment estimation at the server level to stabilize updates. In contrast, the RSF model utilizes a tree-aggregation approach, which requires only a single communication round, significantly reducing the computational and communication overhead compared to iterative gradient-based methods.

• Superiority of RSF: The Random Survival Forest model demonstrated the best overall balance of discrimination, calibration, and robustness across heterogeneous clients, making it the most reliable choice for federated survival analysis in this study.
• Federated vs. Local: Federated learning consistently outperformed local training, proving that collaborative learning effectively mitigates the performance degradation typically seen when models are restricted to small, isolated institutional datasets.