Heterogeneity learning in distributed networks with large-scale survival data

This paper considers survival analysis of large-scale data distributed across heterogeneous network nodes. We propose a novel method, the Distributed Spanning-Tree-Based Fused Lasso (DSTFL), for Cox regression in distributed settings. By employing a minimum spanning tree-based fusion framework, the method can reduce computational and communication burdens, facilitating scalability to large datasets. Additionally, we develop an efficient alternating direction method of multipliers algorithm for the optimization with privacy protection. We establish large-sample properties and clustering consistency for the proposed estimator. Simulation studies demonstrate that DSTFL improves computational efficiency, clustering performance, and robustness compared to existing methods. An application to Surveillance, Epidemiology, and End Results gastric cancer data illustrates how DSTFL identifies geographically structured survival heterogeneity and heterogeneous covariate effects across regions.