Benchmarking Classical, Machine Learning, and Bayesian Survival Models for Clinical Prediction (2509.10073v1)

Abstract: Survival analysis is a statistical framework for modeling time-to-event data, particularly valuable in healthcare for predicting outcomes like patient discharge or recurrence. This study implements and compares several survival models - including Weibull, Weibull AFT, Weibull AFT with Gamma Frailty, Cox Proportional Hazards (CoxPH), Random Survival Forest (RSF), and DeepSurv - using a publicly available breast cancer dataset. This study aims to benchmark classical, machine learning, and Bayesian survival models in terms of their predictive performance, interpretability, and suitability for clinical deployment. The models are evaluated using performance metrics such as the Concordance Index (C-index) and the Root Mean Squared Error (RMSE). DeepSurv showed the highest predictive performance, while interpretable models like RSF and Weibull AFT with Gamma Frailty offered competitive results. We also explored the implementation of statistical models from a Bayesian perspective, including frailty models, due to their ability to properly quantify uncertainty. Notably, frailty models are not readily available in standard survival analysis libraries, necessitating custom implementation. Our results demonstrate that interpretable statistical models, when correctly implemented using parameters that are effectively estimated using a Bayesian approach, can perform competitively with modern black-box models. These findings illustrate the trade-offs between model complexity, interpretability, and predictive power, highlighting the potential of Bayesian survival models in clinical decision-making settings.