Gene expression profiling offers a principled supplement to traditional breast cancer staging, yet most transcriptomic prediction models sacrifice clinical transparency for discriminative performance. We present a biologically informed, explainable machine learning framework that reduces RNA sequencing data to seven curated pathway activity scores and evaluates them through both Cox proportional hazards survival analysis and binary classification, with systematic comparisons against standard pathway scoring methods, molecular subtype baselines, and controlled feature ablation. Seven pathway scores, covering proliferation, estrogen response, immune infiltration, apoptosis, epithelial to mesenchymal transition, HER2 signalling, and angiogenesis, were derived from the TCGA BRCA cohort (n = 213) and validated externally on SCAN B GSE96058 (n = 1,483). Three classifiers (elastic net, random forest, gradient boosting) and a Cox model were compared under stratified five fold cross validation. SHAP analysis produced per patient feature attributions, whose stability was assessed through cross fold rank correlations. Mean z score aggregation was benchmarked against a rank based ssGSEA baseline and a PAM50 molecular subtype only baseline. The combined model achieved AUC 0.856 (random forest) and a Cox concordance index of 0.827, substantially outperforming the subtype only baseline (AUC 0.613; ΔAUC +0.243; ΔC index +0.214). Within Luminal A tumours alone (n = 719), the model separated high risk from low risk patients with a log rank p = 5.9 × 10⁻²² and cross validated C index of 0.848, confirming independent prognostic signal beyond subtype classification. Mean z score aggregation outperformed ssGSEA by 0.010–0.038 AUC across all classifiers. SHAP rankings were stable across folds (Spearman ρ = 0.82) and recovered established oncological mechanisms. Decision curve analysis demonstrated positive net benefit across threshold probabilities 0.04–0.50, and the net reclassification index versus the subtype baseline was 0.801 (95% CI: 0.679–0.911). These results establish that pathway oriented explainable machine learning delivers transparent, generalisable breast cancer risk stratification with prognostic value beyond molecular subtype classification, and that simple mean z score aggregation outperforms the standard ssGSEA approach for compact curated gene sets.
Thayyil et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: