Abstract Summary
This paper tackles the problem of time-to-eventcounterfactual survival prediction, aiming to optimizeindividualized survival outcomes in the presence ofheterogeneity and censored data. We propose CURE, aframework that advances counterfactual survival modelingvia comprehensive multimodal embedding and latent subgroupretrieval. CURE integrates clinical, paraclinical,demographic, and multi-omics information, which are alignedand fused through cross-attention mechanisms. Complexmulti-omics signals can be adaptively refined using amixture-of-experts architecture, emphasizing the mostinformative omics components. Building upon thisrepresentation, CURE implicitly retrieves patient-specificlatent subgroups that capture both baseline survivaldynamics and treatment-dependent variations. Experimentalresults on METABRIC and TCGA-LUAD datasets demonstrate thatproposed CURE model consistently outperforms strongbaselines in survival analysis, evaluated using theTime-dependent Concordance Index (Ctd) and Integrated BrierScore (IBS). These findings highlight the potential of CUREto enhance multimodal understanding and serve as afoundation for future treatment recommendation models.All code and related resources are publicly available tofacilitate the reproducibility.
