Machine learning-based survival modeling reveals distinct surgical benefits across molecular subtypes in young women with de novo metastatic breast cancer

Abstract Background : Survival prediction in de novo metastatic breast cancer (dnMBC) remains challenging due to its heterogeneous molecular subtypes, diverse metastatic patterns, and varied treatment responses. This study aimed to develop accurate prognostic models for overall survival (OS) and cancer-specific survival (CSS) prediction in childbearing-age women (≤ 49 years) diagnosed with dnMBC, utilizing advanced machine learning (ML) techniques and robust statistical methodologies. Methods : A total of 2,605 women diagnosed with dnMBC from 2010 to 2019 were extracted from the SEER database and randomly divided into training (n = 1,833; 70%) and validation (n = 772; 30%) cohorts. Eight ML models—Random Survival Forest (RSF), Gradient Boosting Machine (GBM), Lasso-Cox, CoxBoost, Survival SVM, XGBoost, plsRcox, and superpc—were developed and validated. Model performance was evaluated through area under the receiver operating characteristic curve (AUC), concordance index (C-index), calibration plots, Brier scores, and decision curve analysis (DCA). Inverse probability of treatment weighting (IPTW) was applied to assess the independent survival impact of surgical interventions. Shapley Additive Explanations (SHAP) were employed for interpretability. Results : RSF and GBM demonstrated superior predictive accuracy compared to other ML models. In the validation cohort, RSF achieved the highest C-index (0.789), along with favorable AUC values at 1-year (0.799), 3-year (0.783), and 5-year (0.749) survival intervals, and a low Brier score of 0.134 at 3 years. SHAP analysis consistently identified molecular subtype, metastatic burden, surgical intervention, and chemotherapy as the most influential predictive factors. Surgical intervention significantly improved OS (weighted hazard ratio [HR]: 0.59; 95% CI: 0.52–0.66, p < 0.001) and CSS (weighted hazard ratio [HR]: 0.59; 95% CI: 0.52–0.66, p < 0.001), particularly among HR+/HER2 + and HR−/HER2 + subtypes. Notably, patients with combined bone and liver metastases exhibited superior survival outcomes after breast-conserving surgery compared to mastectomy (median OS: 63 vs. 65 months; both superior to non-surgical management at 37 months; p < 0.001). Conclusions : This ML-based study highlights the prognostic value of molecular subtype, metastatic patterns, and surgical intervention in dnMBC among young women, emphasizing the importance of personalized treatment approaches. Our findings provide robust evidence supporting the integration of ML models in clinical practice, laying a foundation for precision oncology and individualized management of dnMBC. Future prospective trials integrating multi-omics data are warranted to validate these results further.