e13671 Background: Gene expression–based prognostic models and multigene panels are widely used for risk stratification in oncology. However, their performance often degrades when applied across heterogeneous cohorts, cancer types, or underrepresented populations, due to limited target sample sizes and variability in gene–outcome associations. Existing approaches relying on target-only modeling or naive pooling of external datasets may result in instability or bias. We sought to develop a transfer learning framework for genomic time-to-event analysis that enables population-aware prognostic modeling and gene panel recalibration while accounting for cross-cohort heterogeneity. Methods: We developed two complementary transfer learning approaches for high-dimensional survival data. TransSurv performs multi-study genomic survival modeling by selectively borrowing information from auxiliary cohorts using a penalized Cox proportional hazards framework with cohort-level screening to mitigate negative transfer. TransInf focuses on recalibration of existing multigene prognostic panels by integrating target cohort data with compatible external cohorts through gene-level transfer-aware inference. The methods were evaluated using transcriptomic and clinical data from The Cancer Genome Atlas (TCGA), METABRIC, and an independent triple-negative breast cancer (TNBC) cohort from Fudan University Shanghai Cancer Center. Model performance was assessed using concordance index (C-index) and time-dependent area under the curve (AUC) under repeated train–test splits. Results: Across multiple TCGA cancer types and stage-defined subcohorts, TransSurv demonstrated improved discrimination for overall survival compared with target-only penalized Cox models and a state-of-the-art multi-study survival learning approach, with consistent gains in C-index and time-dependent AUC. In the external FUSCC TNBC cohort, TransSurv achieved higher C-index for recurrence-free survival compared with target-only modeling. Using TransInf, recalibrated gene panels derived from established signatures showed improved prognostic discrimination in race-stratified TCGA cohorts and in the TNBC cohort relative to target-only or pooled analyses. The recalibrated panels retained biologically relevant genes while excluding features with inconsistent target-level associations. Conclusions: This transfer learning framework enables robust genomic survival modeling and gene panel recalibration across heterogeneous cohorts by selectively leveraging external data while preserving target-specific inference. The proposed methods support population-aware risk stratification and adaptation of existing prognostic tools, with potential relevance for precision oncology applications in settings with limited target cohort sizes.
Qu et al. (Thu,) studied this question.