Development of machine learning and nomogram models to predict lung metastasis and prognosticate survival in breast cancer

PURPOSE: Lung metastasis in breast cancer (BCLM) is a critical determinant of poor prognosis, occurring in approximately 30-50% of advanced cases and associated with significantly reduced median survival. This study aimed to develop machine learning models for predicting BCLM and evaluating prognosis using the SEER database. METHODS: Data from the SEER database (2018-2021) were analyzed. For the prediction model, 11 independent predictors were identified via univariate and multivariate logistic regression. Machine learning models were developed and evaluated using AUC, accuracy, precision, specificity, recall, F-score. The prognostic model incorporated 12 features through Cox regression, via a nomogram, and validated by C-index, calibration plots, decision curve analysis (DCA), and integrated discrimination improvement (IDI). RESULTS: Among 124,505 patients, 168 (0.135%) developed lung metastasis. Multivariate logistic analysis identified HR-/HER2- subtype (OR = 2.701, 95% CI 1.614-4.52) and brain metastasis (OR = 11.088, 95% CI 3.518-34.946) as independent high-risk factors. The LR-based prediction model demonstrated the highest discriminative ability among the evaluated individual models, achieving an AUC of 0.947 (95% CI 0.902-0.977), sensitivity of 0.816, specificity of 0.911, and an F-score of 0.024. Given the extremely low incidence of lung metastasis (0.135%), the low F-score mainly reflected the limited positive predictive value inherent to rare-event prediction scenarios. An online tool (https://9um39fycfyx4icd6cs8gcw.streamlit.app/) was deployed for risk assessment. 12 factors confirmed by multivariate COX regression were incorporated the nomogram. The prognostic model achieved a C-index of 0.79 (se = 0.009), with 1-year and 3-year survival AUCs of 0.86 and 0.62. The 1-year calibration plots showed high consistency between predicted and observed survival (mean absolute error = 0.001; 0.9 quantile error = 0.003). DCA and IDI confirmed improved clinical net benefits compared to traditional TNM models. CONCLUSION: This study identified key risk factors for BCLM and developed prediction and prognosis models that may assist population-level risk stratification. However, given the rare-event nature of lung metastasis, the prediction model should be interpreted cautiously and is more suitable for risk assessment rather than individual-level screening or diagnostic replacement.