Machine Learning for Breast Cancer Diagnosis with BioGPT-Generated Clinical Explanations

Abstract

Accurate Human Epidermal Growth Factor Receptor 2 (HER2) status assessment is important for breast cancer treatment selection, but machine learning systems for this task must address class imbalance and explainability before they can support clinical decision-making. This paper presents a compact hybrid framework for HER2 status prediction using the METABRIC breast cancer dataset. The framework combines Logistic Regression and XGBoost through stacking, applies class balancing with SMOTE, explains global and patient-level predictions using SHAP, and converts model evidence into clinician- and patient-oriented narratives using BioGPT. To support responsible reporting, the study also highlights minorityclass behavior, baseline fairness, feature-attribution evidence, and limitations related to retrospective data and patient-facing readability. On the original imbalanced dataset, the model achieved 83.3% accuracy but weak HER2-positive detection (F1-score 0.298). After balancing, performance improved to 97.1% accuracy and 0.971 macro F1-score. The proposed model also outperformed eight machine learning baselines, including Random Forest and standalone XGBoost. SHAP analysis identified the most influential variables behind each decision, while BioGPT generated structured narrative explanations from prediction and attribution outputs. The results suggest that combining balanced learning, interpretable feature attribution, and biomedical language generation can improve predictive performance and human-centered explanation, while still requiring leakage-safe validation and external clinical testing before deployment.