Noaman Khan, Zubair Shah
Accurate cancer survival prediction remains challenging, partly because most existing deep learning models offer limited mechanistic interpretability of tumour–immune interactions. We developed MechSurv, a multimodal framework integrating Whole-Slide Images (WSIs) and transcriptomic data through a biologically informed tumour–immune ordinary differential equation (ODE). Patient-specific ODE-parameterised latent features (kinetic proxies) are inferred via learned attention mechanisms, with the ODE serving as a mechanistic feature encoder rather than a system of experimentally measured kinetic parameters. We evaluated the framework on 2238 patients across five cancer types from TCGA using 5-fold cross-validation, benchmarking against a strictly encoder-controlled SurvPath baseline on all five cohorts and against nine established multimodal baselines on breast cancer. MechSurv achieved a C-index of 0.716 ± 0.045 on breast cancer (TCGA-BRCA, n = 868), outperforming an encoder-controlled SurvPath baseline by 3.6 points (0.716 vs. 0.680) with 56% lower cross-validation variance, and showed consistent gains across bladder (0.636), colorectal (0.732), head and neck (0.661), and gastric (0.616) cancers. Pathway–parameter associations were consistent with known tumour–immune biology, and, in an exploratory analysis, ODE-parameterised feature shifts in paired primary–metastatic samples aligned directionally with expected metastatic biology. Structured pruning reduced the pathway-encoder parameters by 62% without performance loss. Integrating mechanistic biological priors into multimodal deep learning can improve survival prediction accuracy, stability, and interpretability; external validation on independent prospective cohorts is required before any clinical application.