Improving Cross-Regional Generalizability of Bridge Deck Deterioration Prediction through Transfer Learning and Explainable Discrete-Time Survival Models

Data-driven deterioration models for bridge decks are trained and validated on pooled national data but are used by individual agencies whose climate, materials and inspection practice differ. On the nationwide National Bridge Inventory panel of 150,144 concrete decks, the published discrete-time gradient-boosted hazard model (DT-GBH) reaches a C-index of 0.769 on a random split but only 0.647 on states withheld from training, no better than Cox regression, while its 5-year calibration error grows from 0.009 to 0.050. We show that this failure combines a shift in the level and shape of the hazard with a shift in covariate effects, and that covariate-shift reweighting and state-blocked model selection do not repair it. We propose TL-DTGBH, which adapts a national DT-GBH to a target state from a short local history in two stages on the hazard-logit scale: a five-parameter, ridge-shrunk recalibration of level, slope and period-specific baseline (HLR), followed by residual boosting whose size is chosen by spell-grouped cross-validation and may be zero. The adapted logit splits exactly into transferred national evidence, a local baseline correction and local covariate corrections, each attributed to features by TreeSHAP. In leave-one-state-out experiments on 14 states, HLR halves the calibration error, from 0.103 to 0.054 with only 50 local sojourns, and the full method raises the C-index from 0.666 to 0.684 with 100 and 0.732 with 500 local sojourns, outperforming local-only and Cox-updating models when local data are scarce and matching a pooled model that must be retrained on national data for every state. Under the original unseen-state protocol, a 20% local history lifts the pooled C-index from 0.646 to 0.757.