Multi-omics survival prediction under modality incompleteness via dynamic modality graphs and product-of-experts fusion

Abstract

Multi-omics data integration provides rich complementary information for cancer survival prediction, yet modality incompleteness is pervasive in real-world medical data and undermines the stability and applicability of existing models. To address this challenge, we propose DMG-PoE, a robust learning framework for multi-omics survival prediction. The framework introduces a dynamic modality graph (DMG) that propagates complementary information across modality nodes to compensate for missing inputs, combined with an uncertainty-aware product-of-experts (PoE) fusion mechanism that adaptively weights each modality by its estimated precision, and a consistency-constrained training strategy that stabilizes predictions across varying modality subsets. Systematic evaluation on the SurvBoard benchmark across 10 TCGA cancer types and seven omics modalities demonstrates that DMG-PoE achieves the highest average C-index under both the standard setting (complete-modality training/testing) and the missing-augmented setting (training additionally includes real incomplete samples and nonclinical modality dropout, with complete-only testing); controlled per-modality masking experiments validate effective test-time recovery in 22/24 cancer-modality combinations, and cross-cohort transfer with only two of seven modalities available further confirms generalization under severe input restriction.