Expert-guided bipolar feature disentanglement for multimodal survival prediction

Multimodal survival prediction holds significant clinical value in precision oncology, increasingly relying on the integration of pathological images and genomic data. However, whole slide images (WSIs) are typically represented as large-scale patches, making them susceptible to noise and regional heterogeneity. Meanwhile, the differences in semantic structure and statistical distribution between modalities lead to entangled intra-modal semantics and hinder effective inter-modal interaction. To address these limitations, we propose a multimodal survival prediction framework based on bipolar disentanglement, termed BPDSurv. The Consensus-Aware Graph Enhancement (CAGE) module constructs spatial and semantic graphs while adaptively suppressing noise through consensus estimation. The BiPolar Fusion (BPF) module disentangles features into positive and negative polarity components and models cross-modal interactions through consistency and complementarity groups, capturing richer survival-related information while preserving intra-modal semantic richness. To support this polarity-specific disentanglement, the Expert-Guided Disentanglement (EGD) module dynamically routes features to the functionally heterogeneous experts, producing complementary representations that enable effective bipolar disentanglement. Extensive experiments on four TCGA cancer datasets demonstrate that BPDSurv achieves leading performance with consistent cross-dataset stability, and comprehensive ablation studies validate the contribution of each component to the complete framework's effectiveness and robustness. Code is available at https://github.com/QAQ404/BPDSurv.