@Article{cmes.2026.079891,
AUTHOR = {Manal Mohammed AL-Tamimi, Siti Norul Huda Sheikh Abdullah, Mohammad Khatim Hasan, Mohammed Azmi Al-Betar, Maw Shin Sim, Abdulrahman Mohammed AL-Tamimi},
TITLE = {Interpretable Cox-Guided Risk Stratification for Specialized Expert Learning in Pan-Cancer Survival Prediction},
JOURNAL = {Computer Modeling in Engineering \& Sciences},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/CMES/online/detail/26782},
ISSN = {1526-1506},
ABSTRACT = {Pan-cancer survival prediction remains a major challenge in personalized oncology due to profound tumor heterogeneity and the complexity of high-dimensional molecular data. Diverse risk profiles across cancer types and noisy, sparse features hinder deep learning models from capturing robust prognostic patterns. Prior pan-cancer studies predominantly focus on multimodal integration or unimodal gene expression analysis, leaving other informative modalities such as Copy Number Variation (CNV) and miRNA expression underexplored. We introduce a new formulation of mixture-of-experts (MoE) survival modeling that recasts expert assignment as a clinically interpretable risk-space decomposition problem. The proposed framework, CoxGuided-SE, constructs an ordered prognostic risk axis from a Cox proportional hazards model trained on clinical covariates, then uses quantile-based thresholds to deterministically stratify patients into risk-homogeneous groups and route them to specialized subnetworks. This design replaces black-box neural routing with a transparent, statistically principled alternative, preventing expert collapse without auxiliary regularization. Evaluated on 33 cancer types from TCGA across four modalities independently, namely clinical features, mRNA expression, miRNA expression, and CNV, CoxGuided-SE achieved substantial gains on high-dimensional genomic data, improving both discrimination and calibration. The most pronounced improvement occurred for CNV, widely considered one of the most challenging modalities for survival modeling, with a 24% increase in discrimination over the strongest baseline. The framework offers two-level interpretability: transparent patient routing based on clinical severity, and expert-specific feature importance within each subgroup. Feature-level analyses further confirm that experts learn complementary and non-redundant representations within each risk stratum. These results provide both a high-performing survival prediction framework and a principled explanation of when and why expert specialization is effective, advancing interpretable and trustworthy clinical AI for pan-cancer prognosis.},
DOI = {10.32604/cmes.2026.079891}
}