An XGBoost–logistic regression hybrid stacking model for breast cancer outcome prediction

Abstract

Accurate prediction of breast cancer outcomes remains challenging due to high-dimensional, imbalanced clinical datasets and the need to balance predictive performance, calibration, interpretability, and model simplicity. We propose an AICc-guided hybrid stacking framework that integrates Logistic Regression and XGBoost using the SEER breast cancer dataset, combining statistical model selection with ensemble learning, explainable artificial intelligence (AI), and calibration-aware evaluation. An information-theoretic model selection approach based on the corrected Akaike Information Criterion (AICc) was incorporated into the logistic regression component to improve parsimony and generalization. Two models were evaluated: Hybrid (Full) and AICc-guided Hybrid (Selected). Performance was assessed using AUC, PR-AUC, F1-score, Brier score, Expected Calibration Error (ECE), Decision Curve Analysis, and a composite Trust Index. Model interpretability was examined using SHAP and LIME. Both hybrid models outperformed standalone Logistic Regression and XGBoost. The AICc-guided Hybrid (Selected) achieved the best overall balance, with an AUC of 0.7176, PR-AUC of 0.3729, F1-score of 0.4078, Brier score of 0.1168, ECE of 0.0294, and the highest Trust Index (0.6704). SHAP and LIME identified clinically meaningful predictors and interaction effects involving tumor size, lymph node involvement, age, and tumor grade. Robustness analysis demonstrated stable performance under perturbations and gradual degradation under increasing noise. The proposed AICc-guided hybrid stacking framework achieves a balanced combination of predictive performance, calibration, robustness, interpretability, and model parsimony, providing reliable and clinically meaningful support for breast cancer risk stratification and decision-making.