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

DMG-PoE delivers the highest average C-index for multi-omics cancer survival prediction on the SurvBoard benchmark across 10 TCGA cancer types, even with modality incompleteness.

• Why it matters: Real-world cancer datasets often miss omics modalities, which can destabilize survival models and limit their use when inputs are incomplete.
• What they did: The method uses a dynamic modality graph to propagate information between modality nodes, an uncertainty-aware product-of-experts fusion that weights modalities by estimated precision, and consistency-constrained training across modality subsets.
• The result: On SurvBoard, DMG-PoE achieves top average C-index in both complete-modality and missing-augmented settings, recovers performance in 22/24 cancer–modality masking combinations, and shows transfer with only 2/7 modalities available.

AI summary of the abstract · may contain errors · not medical advice · read the paper.