Clinical and Genomic Summary Features for Interpretable Basal-Like Breast Cancer Identification: A TCGA-BRCA Machine Learning Study

Abstract

Basal-like breast cancer is an aggressive molecular subtype commonly identified through PAM50 gene-expression profiling. This study examined whether low-dimensional clinical and genomic summary variables could support interpretable basal-like identification. TCGA-BRCA PanCancer Atlas data were obtained from cBioPortal. After excluding 103 patients without PAM50 annotation, 981 patients were analyzed, including 171 basal-like tumors. Logistic Regression and XGBoost were evaluated using clinical-only and clinical-genomic predictors with a stratified 80:20 split, five-fold cross-validation, class weighting, and limited hyperparameter optimization. The clinical-genomic XGBoost model achieved the strongest independent-test performance, with a ROC-AUC of 0.839, PR-AUC of 0.458, sensitivity of 82.4%, specificity of 81.6%, balanced accuracy of 81.9%, and MCC of 0.530. SHAP identified nonsynonymous tumor mutation burden, tumor break load, aneuploidy score, and MSIsensor score as the leading predictors. Results remained directionally consistent after excluding Normal-like tumors. Clinical and genomic summary variables therefore provided moderate and interpretable basal-like identification, although external validation is required before clinical application