Development and external validation of multiple machine learning-based models for breast cancer-specific survival prediction in postoperative patients with invasive breast cancer: a study based on the SEER database and an external cohort.

BACKGROUND

Patients with invasive breast cancer (IBC) make up most breast cancer cases and exhibit significant heterogeneity. Therefore, it is essential to construct an effective model to estimate long-term postoperative breast cancer-specific survival (BCSS).

METHODS

We used data from the Surveillance, Epidemiology, and End Results (SEER) database during 2010-2020, alongside recruiting an external cohort from Guangxi Medical University Cancer Hospital (GMUCH). We constructed four prediction models (Random Survival Forest [RSF], Survival Gradient Boosting Machine [Survival-GBM], Survival Extreme Gradient Boosting [Survival-XGBoost], and Least Absolute Shrinkage and Selection Operator-Cox proportional hazards model [LASSO-Cox]) to predict 3-, 5-, 7-, and 10-year BCSS in patients with IBC.

RESULTS

The RSF model developed in this study exhibited outstanding predictive performance, with a C-index of 0.824 (95% CI: 0.817-0.831) in the training set, 0.689 (95% CI: 0.670-0.707) in the internal validation set, and 0.716 (95% CI: 0.649-0.772) in the external validation set-outperforming its counterparts. Time-dependent Brier scores confirmed the model's excellent calibration and high predictive accuracy. Decision curve analysis (DCA) further confirmed the model's stable clinical utility, while Shapley Additive Explanations (SHAP) diagrams quantified the importance of prognostic features. Additionally, the RSF model demonstrated strong efficacy in stratifying patients into distinct BCSS risk subgroups.

CONCLUSIONS

In conclusion, this study developed an optimal prognostic model for predicting the long-term BCSS of IBC patients, which provides critical support for risk stratification in the clinical management of IBC patients.