Leakage-Aware Explainable Framework for Breast Cancer Prognosis

Abstract

Breast cancer remains a leading cause of cancer-related mortality among women, necessitating reliable and interpretable prognostic models for personalized treatment planning. This study developed a leakage-aware, explainable machine learning framework for breast cancer survival prediction using clinicopathological data from 4,024 patients in the SEER database. The framework integrated feature engineering, leakage detection, hyperparameter optimization, calibration assessment, and explainable artificial intelligence. Among five evaluated algorithms, CatBoost achieved the best performance, with a ROC-AUC of 0.725 and a five-fold cross-validated ROC-AUC of 0.746 (95% CI: 0.708–0.785). SHAP analysis identified Node Ratio, Age, Hormone Index, Tumor Burden, and Grade as the most influential predictors. The proposed framework provides transparent, reliable, and clinically relevant prognostic predictions for breast cancer risk stratification.