HMSP: Hierarchical Multimodal Semantic Fusion of Pathology-Text-Genomic Representations for Robust Survival Prediction

Computational pathology has become an indispensable component of cancer survival analysis, yet state-of-the-art multimodal pipelines still depend heavily on paired whole-slide images (WSIs) and in situ genomic or immunohistochemistry (IHC) profiles. In practice, however, IHC and high-throughput sequencing are costly, tissue-consuming, and frequently unavailable, creating a critical performance gap for cases lacking molecular data. To bridge this gap, we introduce HMSP (Hierarchical Multimodal Survival Prediction), a framework that leverages richly structured pathology reports as a flexible linguistic surrogate for missing genomic information. Concretely, HMSP formulates WSIs, free-text pathology narratives, and (when present) genomic features as a three-level hierarchy of visual, lexical, and molecular semantics. We enforce hierarchical consistency with a Tri-Modal Ranking Contrastive Loss and a Cone-Based Hyperbolic (H-Cone) regulariser, enabling (i) fine-grained alignment between image regions and report phrases and (ii) high-level alignment between report embeddings and genomic biomarkers. Once trained, the report encoder alone can act as a plug-in genomic proxy, allowing robust survival prediction even when molecular assays are absent. Extensive experiments on five public cohorts demonstrate that HMSP not only surpasses previous multimodal baselines under full-data settings, but also maintains state-of-the-art performance when genomic channels are wholly ablated - highlighting the practicality of text-driven surrogates in real-world resource-constrained scenarios. Code and pretrained weights will be released upon publication.