Development and validation of a deep learning model for breast cancer prognosis using the SEER database

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Breast cancer is the most prevalent malignancy and the leading cause of cancer‑related death in women. Despite the widespread use of the TNM stage, patient outcomes vary considerably due to its molecular heterogeneity. Traditional linear models, constrained by linearity assumptions, are inadequate for modeling complex nonlinear interactions in clinical data. Deep learning has emerged as a promising alternative to capture these intricate relationships and improve predictive accuracy. In this study, we developed and validated a DeepSurv‑based model for survival prediction and treatment decision support in patients with breast invasive ductal carcinoma. Methods We identified 110,346 patients diagnosed with breast invasive ductal carcinoma between 2012 and 2015 from the SEER database and randomly divided them into training and testing datasets at a 7:3 ratio. Model performance was evaluated using the C-index, time-dependent ROC, Brier scores, IBS, and DCA. We benchmarked our model against CPH, Lasso, RSF, XGBoost, and clinical staging for comparison. Kaplan-Meier analysis was used to compare survival outcomes. Results DeepSurv model achieved C-index of 0.826 for OS and 0.861 for DSS, outperforming CPH (0.814/0.855), Lasso (0.811/0.854), RSF (0.818/0.858), XGBoost (0.799/0.842), and clinical staging (0.708/0.805). The 3-year and 5-year AUC were 0.868, 0.855 for OS, and 0.901, 0.889 for DSS, with IBS values of 0.142 and 0.118, respectively. DCA curves also confirmed a favorable net clinical benefit. Patients receiving model-concordant treatment showed more favorable OS and DSS than those who did not (both P < 0.001). Conclusion The DeepSurv-based model demonstrated favorable discriminative and calibration performance for individualized survival prediction in breast cancer. While internal validation supports its potential for risk stratification and treatment discussion, external validation in independent cohorts and further interpretability analyses remain essential before clinical translation.