Related Experiment Video
Updated: Oct 22, 2026
Quantifying the Brain Metastatic Tumor Micro-Environment using an Organ-On-A Chip 3D Model, Machine Learning, and Confocal Tomography
Published on: August 16, 2020
Interpretable machine learning model for brain metastasis in breast cancer: a large-scale, multi-center study
Quan Yuan1, Yupeng Sha2, Rui Yu3
1Department of Breast Surgery, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang, 150000, China. yquan646908@gmail.com.
Background:
Brain metastasis (BM) is a devastating complication of breast cancer (BC) with a poor prognosis. Early identification of high-risk patients is essential but currently lacks accurate predictive tools. This study aimed to develop a stable machine learning model for predicting BM in a large, diverse patient population.
Methods:
We conducted a retrospective cohort study of 186,351 primary BC patients from the SEER program, the National Cancer Database, and a multi-center Chinese cohort. Patients were split into training and validation sets. Candidate predictors included demographic, clinicopathological, biomarker, and treatment-related variables, including surgery, chemotherapy, endocrine therapy, neoadjuvant therapy, and radiotherapy. LASSO regression was used for feature selection. Nine machine learning models were developed and compared using area under the curve (AUC) and decision curve analysis. The optimal model was deployed as an online tool.
Results:
Among 186,351 patients, 2,982 (1.6%) developed BM. In baseline comparisons, patients who developed BM more frequently exhibited advanced disease characteristics and different treatment patterns, likely reflecting more aggressive disease presentation. Key predictors retained in the final parsimonious model included AJCC N/M/T stage, age, and surgery status. The LASSO-regularized Logistic Regression model achieved superior discrimination in the validation set, with an AUC of 0.907 (95% CI: 0.857-0.958). The model demonstrated good calibration, clinical net benefit, and interpretability, with SHAP analysis confirming the contribution of each retained predictor to individual risk estimates. Survival analysis confirmed the poor prognosis associated with BM (p < 0.001). An interactive online tool was developed for real-time risk estimation.
Conclusion:
We developed and internally validated a accurate, clinically translatable model for predicting BM in BC using over 186,000 patients from North American and Asian cohorts. This web-based tool can assist clinicians in identifying high-risk individuals, enabling personalized management and potentially improving patient outcomes.
