Federated learning for competing risk analysis in healthcare

Md Mahmudur Rahman, Sanjay Purushotham

Keywords: Survival analysis, Federated Learning, Competing Risks, Censoring

Abstract: Performing survival analysis on distributed healthcare data is an important research problem, as the existing privacy laws and emerging data-sharing regulations prohibit the sharing of sensitive patient data across multiple institutions. The distributed healthcare survival data is typically heterogeneous, non-uniformly censored, and comes from patients with multi-morbidities (or competing risks), which may lead to biased and inaccurate risk predictions. To address these challenges, we propose federated learning for survival analysis with competing risks. Specifically, (a) we propose a simple algorithm to estimate consistent federated pseudo values for survival analysis with competing risks and censoring; and (b) we introduce a novel and flexible federated pseudo-value-based deep learning framework named FedCRA, where we employ a transformer-based model; named TransPseudo, to enable subject-specific prediction of the marginal risk of an event while preserving the data privacy. Extensive experiments on two real-world distributed healthcare datasets with non-IID and non-uniform censoring properties and on synthetic data with different censoring settings demonstrate that our FedCRA framework with the TransPseudo model performs better than the federated learning framework with state-of-the-art survival models for competing risks analysis (CRA).