Horizontal federated learning and assessment of Cox models.

Abstract

The Cox Proportional Hazards model is a widely used method for survival analysis in medical research. However, training an accurate model requires access to a sufficiently large dataset, which is often challenging due to data fragmentation. A potential solution is to combine data from multiple medical institutions, but privacy constraints typically prevent direct data sharing. Federated learning offers a privacy-preserving alternative by allowing multiple parties to collaboratively train a model without exchanging raw data. In this work, we develop algorithms for training Cox models in a federated setting, leveraging survival stacking to facilitate distributed learning. In addition, we introduce a novel secure computation of Schoenfeld residuals, a key diagnostic tool for validating the Cox model. We provide an open-source implementation of our approach and present empirical results that demonstrate the accuracy and benefits of federated Cox regression.