A Multimodal Survival Prediction Fusion Framework Based on Prototype Learning

Accurate survival prediction is critical for cancer prognosis. Multimodal data fusion can improve prediction accuracy, but existing methods face two main challenges: WSIs require partitioning into thousands of patches, increasing computational cost, and genomic data are high-dimensional, sparse, and heterogeneous, hindering cross-modal learning. To address these issues, we propose a multimodal survival prediction framework based on prototype learning and hierarchical fusion. It consists of three key modules: a morphological prototype module that clusters image patches into representative tissue prototypes, a pathway embedding module that maps genomic features into structured pathway space, and a hierarchical fusion module that integrates cross- and self-attention to capture complementary information between pathological and pathway prototypes. Experiments on three TCGA datasets and one private gastric cancer dataset show that our model achieves state-of-the-art performance and can statistically distinguish high- and low-risk patient groups.