Transfer learning with deployment-covariate recalibration for survival prediction under covariate shift

CoxRTL is a new recalibrated transfer-learning framework that borrows information from larger external cohorts to improve Cox-model survival predictions when the target site has limited training data and no deployment-phase outcome labels. The method explicitly addresses covariate shift — the common clinical scenario where patient-mix differences between training and deployment populations degrade model performance. By recalibrating the transferred model to the deployment covariate distribution without requiring outcome data, CoxRTL offers a practical pathway for multi-site oncology prognostic models that must generalise across heterogeneous populations.