Source projection for calibrated transfer learning in survival prediction

Abstract

Transfer learning can improve survival prediction in cohorts with few events. An external model may capture useful covariate associations yet systematically overestimate or underestimate risk in the target population, creating a need for selective borrowing. We propose Source Projection Transfer (SPT) for fixed-horizon prediction with a target generalized additive model. SPT requires only source predictions at target covariates, supporting generalized additive, Cox and neural-network source models without individual-level source data. It projects source log-odds scores onto the target basis and borrows through two separately adapted channels: a projection-centered penalty transfers effect shapes, while intercept-recalibrated soft labels inform probabilities at the target risk level. Under local asymptotics, projection avoids the strictly positive risk cost of estimating an uninformative additional source direction. Each channel asymptotically switches off under its corresponding persistent-mismatch condition. In log-logistic and Weibull simulations with covariate shift, SPT improves mean Brier score and discrimination over target-only modeling in all 24 setting-distribution-horizon combinations. Matched comparisons show improvements over borrowing through source probabilities alone across settings, and over borrowing the complete source score under structural mismatch. In UK Biobank-to-SPRINT mortality prediction among standard-treatment participants with reduced kidney function, SPT improves mean discrimination at 24, 30 and 36 months, with Brier improvements at 24 and 36 months. SPT enables flexible source learners to inform a parsimonious target model while adapting to differences between populations.