ICLR 2026 · 2026

Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis

Yilan Zhang^†, Li Nanbo^†, Changchun Yang, Jürgen Schmidhuber, Xin Gao

† Equal contribution

Paper
Code

Cancer survival modelling must combine information at very different scales, from tissue morphology to genomic activity. SlotSPE represents these modalities as compact sets of prognostic events, allowing the model to reason about distinct patterns and their relationships rather than compressing each modality into one undifferentiated feature vector.

Visualizing prognostic events

SlotSPE histology and genomic slot assignment maps, representative tissue regions, and pathway-level attention scores. — Histology and genomic slot assignments connect learned events to tissue regions and pathway-level attention scores, making their associations inspectable. View full size · Source

From modalities to structured events

Inspired by factorial coding, modality-specific slot attention extracts distinct events from histology images and genomic profiles without event annotations. These slots capture patterns such as tissue structures and pathway coactivations. SlotSPE then models interactions both within and between modalities, incorporating biological priors into the representation.

Evaluation across ten cancer benchmarks examines predictive performance, interpretability, and robustness when genomic information is missing. The event-based representation provides a way to inspect the model’s learned associations alongside its survival predictions.