HyM²Surv: Hypergraph learning for cross-magnification and microbiome-attentive multimodal survival prediction

Survival prediction is a critical task in cancer prognosis assessment. In recent years, deep learning methods based on multimodal data have been widely applied to this task. However, existing studies primarily focus on the integration of whole slide images (WSIs) and genomic data, while the utilization of intratumoral microbiome information remains limited. Meanwhile, current approaches face two limitations: (i) they fail to fully exploit the multi-magnification information of WSIs, limiting the joint representation learning across different resolutions; and (ii) high-magnification pathological features tend to dominate, leading to modality imbalance between pathology and intratumoral microbiome data. In this study, we propose HyM²Surv, a hypergraph-based framework for multi-magnification multimodal survival prediction. Specifically, the model first captures cross-magnification interactions in pathology by leveraging spatial proximity and phenotypic feature similarity. It then constructs dense cross-modal connections by centering on the microbiome and linking it with pathological features across all magnifications, thereby alleviating modality imbalance. Extensive experiments on four TCGA datasets demonstrate that HyM²Surv consistently outperforms state-of-the-art methods, achieving an average improvement of 6.10% in terms of C-index. In addition, the proposed model provides interpretable insights by identifying microbiome-pathology interaction regions that are significantly associated with patient prognosis, highlighting its potential clinical application. The code is publicly available at https://github.com/RanSuLab/HyMMSurv.