JMIR Publications

ABSTRACT

Deep survival models for tumor-tissue and molecular studies are trained only on patients whose specimens are collected, archived, retrieved, assayed, and pass quality control—a covariatedependent draw from the incident-case population that biases the learned relationship toward the analyzable sample rather than the population. We present SBASURV, an architecture-agnostic survival head that replaces the Cox partial likelihood with a stabilized inverse-probability-weighted Cox objective, weighting both the event term and the risk-set denominator, and implements the Lunn–McNeil duplication method as a differentiable module emitting a Wald statistic for etiologic heterogeneity. On five public datasets with induced covariate-dependent selection the head improves held-out concordance over a naive neural head, with weight diagnostics making the bias–variance regime explicit. Deployed on real colorectal cancer across clinical, omics, and whole-slide histopathology with real availability weights, clinical+omics fusion reaches held-out Harrell C 0.68; availability weighting preserves discrimination while shifting the estimand, and a heterogeneity test significant under a mutation-burden proxy subtype is null under a genuine methylation-defined subtype—a caution that subtype conclusions depend on how subtypes are operationalized.