Transportability of Time-Dependent Shapley Explanations of Machine Learning Survival Models in Breast Cancer: A Geographic and Temporal External Validation Study

Background/Objectives: Whether Shapley-based explanations of machine learning survival models are transported to new populations is rarely tested. We assessed the geographic and temporal transportability of the performance and time-dependent SurvSHAP(t) explanations of survival models in breast cancer. Methods: Cox, random survival forest (RSF), gradient-boosted survival (GBSA), DeepSurv and DeepHit models for recurrence-free survival were developed in Rotterdam (n = 2982) and applied without updating to a geographic (German Breast Cancer Study Group [GBSG], n = 686) and a temporal cohort (Duke, 2000–2014, n = 888). Explanation shift was assessed with classifier two-sample tests and density-ratio decomposition; agreement with local reference models refitted in each external cohort was tested with a 500-refit bootstrap null and Holm correction. Results: Discrimination was preserved (Uno’s C at 0.64–0.68 in GBSG; 0.71–0.73 in Duke), but the three-year risk was overestimated about three-fold in Duke (observed/expected at 0.28–0.35); a true decline in risk could not be separated from differences in event ascertainment. Explanations shifted in both cohorts (classifier area under the curve [AUC] at 0.75–0.90), yet nodal category ranked first in all 15 model–cohort pairs. Disagreement with the local reference model was detected only for RSF in GBSG (Holm-adjusted p = 0.032); elsewhere, and particularly in Duke, non-detection does not establish agreement because power was limited. Local models gave more weight to endocrine therapy in Duke (+0.07 to +0.22; +0.04 to +0.24 across imputations). Conclusions: Preserved discrimination did not prevent marked miscalibration in a later cohort, and preserved importance rankings did not guarantee that transported explanations matched those of locally trained models. Explanations should be evaluated externally alongside, and separately from, predictive performance.