Abstract：

Under the background of the Healthy China strategy, in response to the problems such as high misdiagnosis rate, unbalanced data and insufficient interpretability in the diagnosis of benign and malignant breast cancer at grassroots medical institutions, this study, based on the Wisconsin Breast Cancer Database, employed four types of models: XGBoost, Random Forest (RF), LightGBM and Support Vector Machine (SVM), processed the data through data cleaning, standardization, normalization and SMOTE sampling techniques, combined with parameter optimization to construct an auxiliary diagnostic model, and conducted interpretability analysis. The results showed that the XGBoost model performed the best, with an accuracy rate of 0.966 in the test set, an accuracy rate of 0.9388, a recall rate of 0.9583, an F1 score of 0.9485, and an AUC value of 0.9918, significantly superior to other models. Its interpretability analysis clarified the influence of key pathological features such as the average tumor radius and texture mean on the diagnostic results, and the improved data balancing method effectively enhanced the model's ability to identify minority class samples. This study provides high-precision and interpretable technical support for early screening of breast cancer at the grassroots level, and has important reference value for clinical decision-making.

Key Words：

breast cancer; XGBoost; explainable machine learning; auxiliary diagnosis; primary medical care