Explainable machine-learning prediction of overall and cancer-specific survival in adult triple-negative breast cancer using SEER: a comparative study of nomograms and random survival forests.

Background: Triple-negative breast cancer (TNBC) is clinically aggressive and prognostically heterogeneous. Population-based tools that provide transparent, individualized survival estimates and clinically actionable risk stratification for both overall survival (OS) and cancer-specific survival (CSS) remain needed. This study aimed to investigate prognostic factors influencing survival in TNBC patients and to construct and compare the predictive performance of nomogram and random survival forest (RSF) models. Methods: Using the Surveillance, Epidemiology, and End Results (SEER) database, we identified women aged ≥18 years diagnosed with TNBC between 2010 and 2015. After prespecified data cleaning, 27,256 patients were included and randomly split into training (70%, n=19,079) and test (30%, n=8,177) cohorts. For OS and CSS, we developed multivariable Cox proportional hazards models and constructed nomograms to predict 1-, 3-, and 5-year survival probabilities. In parallel, RSF models were trained under the same split, and SHapley Additive exPlanations (SHAP) were used to interpret RSF predictions. Model performance was evaluated using Harrell's concordance index [C-index; bootstrap 95% confidence intervals (CIs)] and time-dependent receiver operating characteristic (ROC) curves with area under the curve (AUC) at 1, 3, and 5 years; calibration was assessed by time-specific calibration plots; clinical utility was examined by decision curve analysis (DCA); and risk stratification was tested using Kaplan-Meier curves with a training-derived median cutoff applied unchanged to the test cohort. Results: Baseline characteristics were well balanced between the training and test cohorts (all P>0.05). In multivariable Cox analyses, older age, higher nodal stage, and metastasis-related variables were independently associated with worse OS and CSS, whereas surgery, radiotherapy, and chemotherapy were associated with lower hazards after adjustment. The nomograms demonstrated consistent discrimination and calibration. RSF showed overall superior discriminative performance compared with the Cox-based nomograms, most evidently in the training cohort, and maintained competitive performance in the test cohort. SHAP consistently highlighted nodal stage, T stage, and age as leading contributors to RSF predictions. Both nomogram- and RSF-based risk scores yielded robust separation of high- versus low-risk groups in both cohorts (all log-rank P<0.001). DCA indicated that both models achieved higher net benefit than treat-all and treat-none strategies across a broad range of threshold probabilities at 1, 3, and 5 years, with RSF generally matching or exceeding the nomogram across much of the threshold range. Conclusions: In this large SEER-based TNBC cohort, RSF models delivered overall superior discrimination and comparable or better decision-analytic net benefit relative to Cox nomograms, while SHAP provided transparent attribution of key predictors. These models support individualized prognosis estimation and clinically meaningful risk stratification for both OS and CSS within an internal validation framework.