Disentangling Consistency, Complementarity, and Noise for Multimodal Cancer Survival Prediction
Abstract
Integrating heterogeneous cancer data, such as histopathological images and genomic profiles, provides a multimodal perspective for assessing patient survival and improves predictive performance. However, existing methods often overlook the complex entanglement among multimodal shared prognostic information (consistency), modality-specific prognostic information (complementarity), and task-irrelevant information (noise), which substantially hinders progress in cancer survival prediction. To address this challenge, we propose a novel multimodal disentangled representation learning framework that effectively separates consistency, complementarity, and noise, thereby mitigating modality heterogeneity and noise interference. Specifically, we propose a task-driven disentanglement module built on a Transformer architecture to decompose multimodal features into consistency, complementarity, and noise representations, while enforcing their distance relationships via multiple regularization losses. To further capture complex intra-modal and cross-modal interactions, we design a disentangled representation graph fusion strategy that constructs consistency–complementarity graphs and aggregates disentangled representations via graph attention. Quantitative analyses and visual evaluations on five cancer datasets consistently demonstrate the superiority of our method over existing survival prediction methods, highlighting its strong capability to exploit rich survival-related information from multimodal cancer data.