Learning Invariant Graph Representations for Cox Survival Modeling under Distribution Shifts

Abstract

Survival prediction from high-dimensional biomedical data is frequently compromised by distribution shifts across multi-center cohorts, where models trained on specific populations often rely on spurious correlations that fail to generalize to new environments. While recent independence-driven reweighting techniques attempt to mitigate this, they typically treat patients as isolated instances, neglecting the intrinsic topological structures and biological pathways shared within patient populations. To address this limitation, we propose InvGraphCox (Invariant Graph Cox), a novel framework that integrates graph-structured representation learning with robust survival modeling. InvGraphCox constructs a k-nearest-neighbor patient graph to capture local manifold structures and employs a Variational Graph Autoencoder (VGAE) combined with a cohort-wise alignment mechanism to learn low-dimensional patient embeddings that are invariant to site-specific biases. We comprehensively evaluate the framework across three distinct experimental settings: the Curated Top-100 Gene Benchmark for stable biomarker identification, large-scale, high-dimensional transcriptomic datasets (Ovarian and Breast Cancer) for unsupervised representation learning, and clinical datasets (Breast and Lung Cancer) involving mixed-type covariates. Experimental results demonstrate that InvGraphCox consistently outperforms state-of-the-art baselines in terms of discrimination, calibration, and risk stratification, confirming its ability to extract robust, biologically meaningful representations in heterogeneous healthcare settings.