A Comparative Evaluation of Traditional, Machine Learning, and Deep Learning Survival Models for Breast Cancer Survival Prediction.

Abstract

Breast Cancer is the leading cause of death from cancer in women worldwide. In our study, we compared the Cox proportional hazards model, Random Survival Forest, Survival Support Vector Machine, DeepSurv, and XGBoost Survival within a unified analytical framework. The analysis used 1,981 patients with complete overall survival information from the METABRIC dataset, including 1,144 deaths and 837 censored observations. Five-fold cross-validation was used for Cox PH and Random Survival Forest, while 10-fold cross-validation was used for Survival SVM, DeepSurv, and XGBoost Survival. Model performance was evaluated using Harrell’s C-index, Uno’s C-index, time-dependent AUC, Integrated Brier Score, calibration analysis, decision curve analysis, and explainability methods. Survival SVM achieved the highest Harrell’s C-index of 0.7011 and AUC of 0.7541, whereas Random Survival Forest achieved a comparable C-index of 0.6970 and the lowest IBS of 0.1725. Random Survival Forest showed the most favorable prediction error, calibration, and clinical net benefit among models that generated survival probabilities. These findings indicate that Survival SVM provides the strongest discrimination, while Random Survival Forest offers the best overall balance of discrimination, probability accuracy, interpretability, and potential clinical utility for breast cancer survival prediction. Explainability analyses consistently identified age at diagnosis, Nottingham Prognostic Index, tumor size, and positive lymph-node involvement as the most influential predictors of mortality risk.

Keywords: Breast cancer, Random survival Forest, Cox Proportional Model, DeepSurv, Survival Support Vector Machine, XGBoost Survival, Kaplan–Meier Estimator.

DOI: 10.7176/CEIS/17-1-08

Publication date: September 28^th, 2026