Federated Survival Analysis Evaluated on Heterogeneous Breast Cancer Data.

Survival analysis is a cornerstone of clinical decision-making, yet building robust time-to-event models often requires extensive, diverse patient cohorts. Such data is rarely available at a single institution, and privacy regulations prohibit centralizing sensitive patient information. Federated learning (FL) offers a privacy-preserving solution by enabling collaborative model training without sharing raw data. This research presents a thorough multi-model evaluation of federated survival analysis, specifically using a cross-institutional breast cancer cohort with naturally heterogeneous distributed clients. The study compared three prominent survival models—Cox Proportional Hazards, DeepSurv, and Random Survival Forest (RSF)—across centralized, local, and federated training paradigms. Additionally, three federated optimization strategies (FedAvg, FedProx, and FedAdam) were assessed for gradient-based models. The findings indicate that FL consistently surpasses local training performance and frequently achieves or even exceeds the performance of centralized models. Among the evaluated models, Random Survival Forest (RSF) emerged as the most balanced in terms of discrimination, calibration, and robustness across diverse client distributions. The study also highlighted that performance is influenced by client data diversity, with FedAvg and FedProx proving more stable and effective than FedAdam. These results provide practical guidelines for selecting models and training paradigms for federated survival modeling in healthcare.