Machine learning-based prognostic model for metastatic breast cancer and its interpretability: a multicenter retrospective study.

Background: Prognostic evaluation of metastatic breast cancer (MBC) currently confronts a two-fold challenge: suboptimal accuracy of conventional scoring systems and insufficient clinical interpretability of machine learning models. This study aimed to construct and validate an accurate prognostic model for predicting the overall survival (OS) of patients with MBC in the Surveillance, Epidemiology, and End Results (SEER) database using machine learning (ML) techniques. Methods: A total of 1,385 MBC patients were enrolled from the SEER database and randomly assigned into the training cohort (1,035 cases) and the internal validation cohort (350 cases). An external validation cohort comprising 73 patients from Jiaxing Women and Children's Hospital was also set up. The key characteristics influencing the OS were identified through multivariate Cox regression analysis, and prognostic models were constructed using four ML algorithms. Results: The random survival forest (RSF) model achieved the best performance both in the training and internal validation cohorts, with a concordance index (C-index) of 0.723 [95% confidence interval (CI): 0.704-0.740] and 0.727 (95% CI: 0.693-0.761), respectively. Notably, the area under the curve and Brier scores of the RSF model exceeded those of other models, confirming its superior survival prediction performance. The decision curve analysis (DCA) further indicated that the RSF model could effectively predict the 1-, 3-, and 5-year OS, making it ideal for clinical application. In the external validation cohort, the C-index of the RSF model was 0.685 (95% CI: 0.606-0.758), which, although slightly lower compared with that recorded in the training cohort, was more stable. The area under the curve and Brier scores further confirmed high accuracy and calibration power of the model. The SHapley Additive exPlanations (SHAP) analysis revealed that triple-negative breast cancer (TNBC) and brain metastasis were core variables that increased mortality risk. Conclusions: The constructed RSF prognostic model demonstrated excellent predictive performance in MBC survival prediction and achieved good interpretability as confirmed by the SHAP analysis. These findings indicate that the developed model can facilitate prognostic assessment and promote the design of individualized treatments for MBC patients.