Machine learning-based prognosis and early death prediction in de novo stage IV breast cancer patients with bone metastasis: a SEER database and multicentre retrospective study

Bone is the most common site of distant metastasis in breast cancer (BC), and the development of bone metastasis (BM) is associated with reduced survival. However, reliable clinical models for accurately predicting prognosis in these patients are lacking. Furthermore, the therapeutic benefit of primary site surgery in patients with de novo stage IV breast cancer (DnIV BC) and BM remains controversial. We extracted data for 23,723 patients diagnosed with DnIV BC and BM between 2010 and 2020 from the Surveillance, Epidemiology, and End Results (SEER) database. Using this data, we developed and compared multiple machine learning models to predict survival and the risk of early death. The models were subsequently validated on an independent, multi-center Chinese cohort of 230 patients. We also performed Kaplan-Meier survival analyses before and after propensity score matching to assess the impact of primary tumor surgery. The Light Gradient Boosting Machine (LightGBM) model exhibited promising predictive performance across multiple evaluation settings. On the test dataset, the model's area under the curve (AUC) of the receiver operating characteristic (ROC) for 12-, 36-, and 60-month survival was 0.775, 0.745, and 0.730, respectively. External validation demonstrated moderate discrimination but reduced performance (AUC drop to 0.67), indicating limited generalizability. In a sensitivity analysis excluding post-diagnostic treatment variables, the model achieved an AUC of 0.768 for early death prediction. External validation of the early death model was not feasible because of the limited number of early death events in the Chinese cohort. Survival analyses showed that primary tumor surgery was associated with improved overall survival after propensity score matching. However, because treatment response and performance status were unavailable, residual confounding and selection bias cannot be excluded. The LightGBM model provides a potentially useful tool for prognostic prediction in DnIV BC patients with BM. Primary tumor surgery was associated with favorable survival outcomes, although causal inference is limited by the retrospective observational design and potential residual confounding.