A Biologically Informed Explainable Machine Learning Framework for Breast Cancer Progression Risk Stratification From Tumor Gene Expression

This study presents a biologically informed explainable machine learning framework for breast cancer progression risk stratification using tumor gene expression data. We engineer pathway-level features from curated gene sets spanning proliferation, estrogen response, immune activation, invasion/EMT, apoptosis, HER2 signaling, and angiogenesis. Models are trained on TCGA (n=213) and externally validated on GSE96058 (n=1,483). We integrate Cox proportional hazards survival analysis (C-index 0.827), ssGSEA baseline comparison, SHAP-based feature importance with cross-fold stability analysis (mean Spearman ρ=0.820), and calibration assessment. The framework achieves AUC 0.856 on external validation with patient-level interpretability through SHAP waterfall plots.