A Calibrated Stacking Ensemble Framework for High-Fidelity Breast Cancer Mortality Prediction Ensemble for Cancer Prognosis

Abstract

Background: Accurate breast cancer mortality prediction is constrained by severe class imbalance and the sensitivity-specificity trade-off, which frequently induces high false-positive rates in clinical settings. To address this limitation, we propose a rigorously calibrated Static Stacking Ensemble framework.
Methods: The model was developed from the METABRIC cohort (N=1816) using 5-fold stratified cross-validation, while the SEER dataset (N=4024) was reserved for independent external validation. The architecture combines out-of-fold (OOF) predictions from deep neural network, XGBoost, and random forest base learners. To replace the standard fixed decision threshold (default τ=0.5), we used a data-driven threshold based on Youden’s J statistic.
Results: Baseline models had important limitations, including a specificity of 0.1997 in standalone deep learning and weight collapse in evolutionary ensembles. At the optimized decision threshold (τ=0.7378), the proposed model balanced sensitivity and specificity. Evaluated on the independent cohort, the model yielded a specificity of 0.6996, a sensitivity of 0.7298, an accuracy of 0.7212, and an F1 score of 0.7894. The framework achieved a ROC-AUC of 0.6925, an area under the precision-recall curve (PR-AUC) of 0.8914, and a Brier score of 0.1563.
Conclusion: Combining OOF prediction stacking with an optimized decision threshold provides a calibrated clinical decision support system (CDSS). This framework minimizes prognostic false alarms and robustly stratifies high-risk patients across heterogeneous cohorts, significantly outperforming isolated baseline architectures.