External Validation and Calibration of Hybrid Deep–Boosted Ensembles for Breast Cancer Survival Prediction: A Cross-Cohort Study on SEER and METABRIC

Abstract Purpose Hybrid ensembles that combine deep neural networks with gradient-boosted tree learners are often reported to improve clinical prediction, but their benefit under leakage-free evaluation remains unclear. Methods We evaluated seven model families—logistic regression (LR), random forest, XGBoost, LightGBM, multilayer perceptron (MLP), MLP + XGB, and MLP+LGBM—for breast cancer survival-status prediction using SEER (n = 4,024; 15.2% mortality) with external validation on METABRIC (n = 1,904; 57.9% mortality). The protocol used repeated nested cross-validation with per-fold selection of imbalance-handling strategy and hybrid blending weight, multi-seed reruns, Nadeau–Bengio corrected paired t-tests, Holm–Bonferroni adjustment across 63 comparisons, and paired bootstrap confidence intervals. Results No hybrid ensemble differed significantly from any single model on F1 (Dead) or macro-F1 after correction. The hybrid blending weight varied substantially across folds. On METABRIC, LR achieved the highest F1 (Dead) on both imputed (0.636) and complete-case (0.641) sets, while calibration analysis showed that the MLP Brier score more than doubled between cohorts. Conclusion For the evaluated structured feature space and validation protocol, hybrid deep–boosted ensembles did not provide robust gains over a simpler class-weighted logistic-regression baseline. Cross-cohort degradation was concentrated in calibration rather than discrimination, suggesting that target-cohort recalibration may be more impactful than source-cohort model complexity.