Design and evaluation guidelines for survival models in federated healthcare settings

Survival analysis is widely used in healthcare to model time-to-event outcomes such as disease progression or patient mortality. However, its application is often limited by the decentralized and privacy-sensitive nature of medical data, which is distributed across multiple institutions. This thesis explores federated learning (FL) as a privacy-preserving approach for collaborative survival modeling without requiring data to be centralized. The performance of Cox Proportional Hazards (CoxPH), DeepSurv, and Random Survival Forests (RSF) is evaluated under local, centralized, and federated training paradigms using the Fed-TCGA-BRCA dataset, which reflects heterogeneous, multi-institutional healthcare data. Model performance is assessed using the concordance index (C-index), time-dependent area under the curve (AUC), and Integrated Brier Score (IBS). Results show that centralized training achieves the highest discrimination, with C-index values of approximately 0.73, compared to 0.61 and 0.60 for federated and local CoxPH models, respectively. However, FL consistently improves over local training and achieves performance close to centralized approaches, particularly for more flexible models. For instance, DeepSurv achieves a C-index of approximately 0.73 under federated training, outperforming centralized (0.71) and local (0.62) settings. In terms of calibration, FL provides lower IBS values, such as 0.15 compared to 0.16 for local training, indicating more reliable survival probability estimates. Across models, RSF and DeepSurv outperform CoxPH, with RSF achieving the lowest IBS values (0.138) and strong robustness to data heterogeneity. These results demonstrate that FL is a viable alternative when data sharing is restricted, balancing privacy and predictive performance, and provide practical guidelines for selecting survival models and training strategies under varying data conditions.