To develop and validate a machine learning-based prognostic model specifically for triple-negative breast cancer (TNBC) patients with axillary lymph node metastasis (ALNM).
Approach:
Study Design: Retrospective analysis of 19,289 TNBC patients with ALNM from the SEER database (2015-2020), divided into training (n=13,502) and validation (n=5,787) cohorts.
Model Development: Five machine learning survival models were developed and compared: Cox Proportional Hazards, Random Survival Forest, Extremely Randomized Survival Trees, Gradient Boosting Survival Analysis, and Survival Tree.
Performance Evaluation: Model performance was assessed using C-index, time-dependent AUC, Brier scores, calibration curves, and decision curve analysis.
Interpretability Analysis: SHapley Additive exPlanations (SHAP) analysis was utilized to identify key prognostic drivers.
Key Findings:
Thirteen independent prognostic factors were identified, including demographic characteristics such as age, race, marital status, and household income; tumor features including histology type, T stage, N stage, M stage, tumor grade, and tumor size; and treatment modalities such as surgery, radiotherapy, and chemotherapy.
The ensemble models (ERST, RSF, GBSA, CoxPH) showed comparable C-indices, with ERST selected for SHAP interpretation.
Time-dependent AUC values indicated excellent discriminatory ability for 1-, 3-, and 5-year survival predictions.
SHAP analysis identified tumor grade, N stage, and radiotherapy as the most influential prognostic factors.
Interpretation:
The ERST model demonstrated robust performance, addressing a critical gap in precision oncology for TNBC patients with ALNM.
Limitations:
The study is based on retrospective data, which may limit generalizability.
Further prospective validation in independent cohorts is necessary.
Conclusion:
The study presents the first machine learning-based prognostic model for TNBC patients with ALNM, facilitating individualized risk assessment and clinical translation.
