Novel models by machine learning and algorithm improvement to predict the risk of breast cancer lung metastasis

LIU Siqi; ZHOU Yanlong; LIN Fangcai; SUN Xin; LIU Na; YU Hongwei

doi:10.3969/j.issn.1009-0460.2025.08.007

Abstract：Objective To build a breast cancer lung metastasis risk prediction model based on 9 machine learning algorithms.Based on the optimal performance model,the algorithm is optimized to further improve the prediction effect of the model,and finally the visual risk assessment tool is constructed.Methods Patients in the Surveillance,Epidemiology,and End Results(SEER)database were screened according to inclusion criteria and exclusion criteria.Logistic regression model and least absolute shrinkage and selection operator(LASSO)regression analysis were used to select clinical features.Based on decision tree(DT),Logistic regression(LR),random forest(RF),K-nearest neighbor(KNN),support vector machine(SVM),naive bayes(NB),extreme gradient boosting(XGBoost),stochastic gradient boosting tree(SGBT),artificial neural network(ANN),9 machine learning algorithms were used to construct the prediction model.Through the introduction of cost sensitive learning,the optimal predictive performance of the model was improved,and the breast cancer lung metastasis risk calculator was established.Results A total of 11 166 breast cancer patients were included in this study,and 15 characteristics with statistical significance were selected.SGBT with the best predictive performance(AUC=0.717)was selected for subsequent algorithm improvement,and the results showed that the improved model was the best comprehensive predictive performance of CS-SGBT(accuracy rate:0.713,recall rate:0.710,F-measure:0.602,AUC:0.788).SHAP analysis results showed that the contribution order of the 15 characteristics was as follows:distant lymph node metastasis,bone metastasis,T stage,surgery,grade,N stage,tumor site,radiotherapy,age,marital status,subtype,brain metastasis,race,estrogen receptor expression and sex.Finally,the risk calculator of breast cancer lung metastasis was constructed based on CS-SGBT and verified by 63 advanced breast cancer patients in Beijing Electric Power Hospital(accuracy rate:0.639,recall rate:0.406,F-measure:0.542,AUC:0.798).Conclusion In this study,an interpretable CS-SGBT-based prediction model was constructed through machine learning and algorithm improvement,which provides a good reference value for clinically assessing the risk of breast cancer lung metastasis.