23 September 2026 in SSRN Electronic Journal
Background Pan-cancer survival models trained on TCGA gene expression routinely report areas under the receiver-operating curve (AUC) of 0.80 to 0.85. We investigate the extent to which these apparently strong discrimination values reflect true within-cancer prognostic performance. Methods Using overall-survival labels derived from confirmed clinical endpoints, a leakage-aware pipeline and four model families evaluated out-of-sample, we reproduce the familiar result: logistic regression attains a pooled test AUC of 0.828 and a Cox elastic-net model a concordance index of 0.749. Results We then examine the factors contributing to this predictive performance. A clinical baseline built from routine chart variables reaches AUC 0.801, and tumor type alone reaches 0.779; adding the year of diagnosis, a purely administrative variable, contributes a further 0.014. Expression improves on the clinical chart by a significant but small margin of +0.027 AUC, whereas clinical variables provide no measurable incremental discrimination once expression is included, because the selected components already recover tumor type (95.6% accuracy), patient sex (AUC 0.981), and contributing institution (62.7% accuracy), consistent with a batch effect. Critically, when evaluated within tumor type rather than pooled, the expression AUC falls from 0.828 to 0.693. We further show that the censoring-based exclusion is not random: excluded patients differ from those retained by diagnosis year, stage, and site. We characterize where the transcriptome adds value within diagnosis, report a sparse biomarker panel of which four genes survive correction in held-out data, and show that a common attribution workflow is unstable across analyses. Conclusions The apparent accuracy of pan-cancer survival models is therefore substantially an artefact of tumor-type composition and cohort selection rather than within-cancer prognostic discrimination