Evolving Sparse P2P Networks For Federated Survival Prediction With Adaptive Utility

Abstract

Federated survival analysis in healthcare must contend with heterogeneous data, right-censoring, and evolving patient populations. While existing fixed client-server federated learning (FL) architectures limit adaptability and create coordination bottlenecks, we envision a peer-to-peer (P2P) framework in which clients adaptively select neighbors to maximize predictive utility, balancing survival accuracy under concept drift at the center level with sparse network construction to control communication cost. Specifically, we apply two objectives to inform P2P network topology evolution, namely, a social-welfare formulation optimizing the collective network benefit and a pareto-efficient local-utility formulation, where no single peer can improve its own utility without undermining the utility of at least one neighbor. Repeated, gain-driven link reassessment combined with a hierarchical time-scale of inner communication rounds and outer population epochs enables the network to adapt to distributional drift while preserving privacy via differentially private parameter sharing. Rigorous experiments on synthetic and real-world SEER data show that our evolving P2P framework achieves centralizedlevel accuracy while reducing communication overhead by over 95% in non-IID settings in comparison with existing baselines.

Department(s)

Computer Science

Keywords and Phrases

Dynamic Networks; Federated Learning; Graphical Game Theory; Healthcare; Pareto Efficiency; Peer-to-Peer Networks; Survival Analysis; Topology Optimization

Document Type

Article - Conference proceedings

Document Version

Citation

File Type

text

Language(s)

English

Rights

© 2026 Institute of Electrical and Electronics Engineers, All rights reserved.

Publication Date

01 Jan 2026

Link to Full Text

Share

 
COinS