Machine learning-based prediction of survival in breast cancer patients with lung metastasis: a SEER-based study.

BACKGROUND: Breast cancer remains the most prevalent malignancy among women, and breast cancer with lung metastasis (BCLM) significantly increases morbidity and mortality. Accurate prognosis prediction for BCLM patients is critical for guiding personalized treatment strategies, yet effective predictive tools remain limited. This study aimed to develop a machine learning (ML)-based predictive framework for BCLM prognosis by leveraging data from the Surveillance, Epidemiology, and End Results (SEER) database (2011-2020). METHODS: We analyzed data of 2,921 female BCLM patients from the SEER database (2011-2020). Advanced oversampling techniques (SMOTEENN, SMOTETomek) addressed class imbalance, and Bayesian optimization tuned model hyperparameters. Six ML classifiers-linear regression, elastic net, random forest, support vector machine (SVM), extreme gradient boosting (XGBoost), and gradient boosting machine (GBM)-were evaluated. SHapley Additive exPlanations (SHAP) provided model interpretability and identified key prognostic factors. RESULTS: SVM achieved the highest performance in 1-year prediction [area under the curve (AUC): 0.770; 95% confidence interval (CI): 0.737-0.803], while XGBoost demonstrated slightly higher discrimination in 3-year (AUC: 0.770; 95% CI: 0.739-0.802) and 5-year (AUC: 0.746; 95% CI: 0.709-0.783) predictions. Key prognostic factors-marital status [specifically for the hormone receptor (HR)+/human epidermal growth factor receptor 2 (HER2)- subtype] and metastatic sites (bone/liver)-were consistently identified as critical determinants of survival across models. SHAP analysis revealed interactions between sociodemographic and molecular factors. CONCLUSIONS: The findings demonstrate that ML-based frameworks, particularly those integrating multimodal clinical data, can improve risk stratification for BCLM patients and provide actionable insights for personalized oncology. Limitations include reliance on retrospective registry data and unmeasured confounders; future research should incorporate longitudinal and imaging biomarkers to further refine predictive accuracy.