Transportability Of Pathway-Based Survival Models From RNA Sequencing To Microarrays In Gastric Cancer: A Multi-Cohort Study

Whether pathway-based survival models developed from RNA sequencing preserve patient risk ordering in external microarray cohorts remains uncertain. Differences in assay platform and patient-cohort composition jointly shape external performance. We quantified joint cross-platform and cross-cohort transportability and compared modelling strategies within a leakage-controlled benchmark. Three pathway encodings, two featureselection states, and five survival learners formed 30 systems developed in TCGA-STAD. The development cohort included 387 patients; four external cohorts included 975 patients. Before external outcome evaluation, the protocol-locked analysis fixed the encodings, feature states, learner set, four GEO cohorts, fitted parameters, and prediction files. Regularised Cox provided the most consistent external discrimination, with an unweighted four-cohort macro Harrell C-index of 0.582, indicating modest preservation of risk ordering. Graph-based and recent neural learners varied across cohorts. Stability selection produced the same external patient ordering as its matched identity-gate Cox model. A separately frozen clinical–molecular model was evaluated in GSE183136 after removal of overlapping expression profiles and failed both prespecified confirmation criteria. This study defines a controlled pipeline for evaluating transcriptomic prognosis across assay platforms and patient cohorts. Its results support cohort-level uncertainty reporting, clinical-reference comparison, and prospectively frozen external evaluation in gastric cancer computational oncology.