Machine learning-based prognostic model for triple-negative breast cancer with axillary lymph node metastasis.

Background: Triple-negative breast cancer (TNBC) with axillary lymph node metastasis (ALNM) represents a high-risk population with substantially worse prognosis compared to other breast cancer subtypes. Despite the critical clinical importance of accurate prognostic assessment in this population, no validated risk prediction model specifically tailored for TNBC patients with ALNM currently exists. Machine learning approaches offer the potential to integrate multiple clinical and pathological factors for improved risk stratification, yet their application in this specific context remains unexplored. Methods: This retrospective study analyzed 19,289 TNBC patients with ALNM from the Surveillance, Epidemiology, and End Results (SEER) database (2015-2020). Patients were randomly allocated to training (n=13,502, 70%) and validation (n=5,787, 30%) cohorts. Independent prognostic factors were identified through univariable and multivariable Cox regression analysis. Five machine learning-based survival models were developed and compared: Cox Proportional Hazards (CoxPH), Random Survival Forest (RSF), Extremely Randomized Survival Trees (ERST), Gradient Boosting Survival Analysis (GBSA), and Survival Tree (ST). Model performance was evaluated using concordance index (C-index), time-dependent area under the curve (AUC), Brier scores, calibration curves, and decision curve analysis. SHapley Additive exPlanations (SHAP) analysis was employed to enhance model interpretability and identify key prognostic drivers. Results: Multivariable Cox regression identified 13 independent prognostic factors encompassing demographic characteristics (age, race, marital status, household income), tumor pathological features (histology type, T stage, N stage, M stage, tumor grade, tumor size), and treatment modalities (surgery, radiotherapy, chemotherapy). The four ensemble/regression models (ERST, RSF, GBSA, CoxPH) demonstrated comparable C-indices (0.7494, 0.7489, 0.7483, and 0.7455, respectively) and substantially outperformed ST (0.6959); ERST was selected for downstream SHAP interpretation given its marginally highest C-index. Time-dependent AUC values for 1-, 3-, and 5-year survival predictions were 0.889, 0.773, and 0.740, respectively, with corresponding Brier scores of 0.014, 0.067, and 0.151, indicating excellent discriminatory ability and calibration. Decision curve analysis confirmed favorable clinical utility across a wide range of threshold probabilities. SHAP analysis revealed tumor grade, N stage, and radiotherapy as the three most influential prognostic factors, with high tumor grade, advanced nodal stage, and absence of radiotherapy consistently associated with increased mortality risk. Conclusions: We developed and internally validated the first machine learning-based prognostic model specifically for TNBC patients with ALNM, integrating 13 clinicopathological variables. The ERST model demonstrated robust discriminatory performance, excellent calibration, and favorable clinical utility. SHAP-based interpretability analysis provided transparent insights into key prognostic drivers, facilitating individualized risk assessment and clinical translation. This tool addresses a critical gap in precision oncology for this high-risk population and has the potential to inform treatment decisions, optimize surveillance strategies, and improve prognostic counseling. Future prospective validation in independent cohorts and integration of molecular biomarkers represent important next steps toward clinical implementation.