Personalized Biomarker Prediction in Breast Cancer Brain Metastases via Ensemble Learning Integrating Radiomics- Clinical Integration

다국어 초록 (Multilingual Abstract)

In breast cancer brain metastases (BCBM), the accurate evaluation of the estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) is essential for personalized treatment. However, tissue from brain lesions can only be obtained through invasive and risky stereotactic or surgical biopsy, which is rarely done in routine care. To address this, we developed a radiomics-based ensemble machine-learning model for the noninvasive prediction of ER, PR, and HER2 status. We leveraged The Cancer Imaging Archive BCBM-Radio Genomics dataset and retrospectively analyzed 165 patients with histologically confirmed BCBM by extracting radiomic features from T1-weighted post-contrast magnetic resonance imaging. After preprocessing the data with ADASYN for class balancing, correlation-based feature reduction, and Isolation Forest for outlier detection, we built an ensemble model that incorporated six algorithms (Random Forest, Extra Trees, Balanced Random Forest, XGBoost, LightGBM, and CatBoost) via ensemble voting. The resulting model achieved test accuracies of 82.1%, 83.0%, and 83.4% for ER, PR, and HER2, respectively, demonstrating its strong potential for noninvasive molecular prediction in BCBM.