Simulation study compares Cox model and machine learning for survival prediction

Researchers have published a comprehensive simulation study in BMC Medical Research Methodology that evaluates the predictive performance of the traditional Cox proportional hazards model against several machine learning methods for time-to-event data. The study, authored by Josline Adhiambo Otieno, Jenny Häggström, and Marie Eriksson, addresses a persistent challenge in medical data analysis: determining which statistical model is most suitable for accurate predictions in specific clinical applications. By utilizing simulation, the team was able to conduct a fair and comprehensive comparison of overall survival prediction performance using both simulated and real datasets, ensuring that the models were tested under equal-sized data conditions.

The primary motivation for this research stems from the proliferation of data-driven risk prediction models designed for analyzing time-to-event data. While many such models have been developed, selecting the most appropriate one for a given medical context remains difficult. The authors argue that simulation provides an effective mechanism for comparison because it allows for controlled conditions where dataset sizes are equalized.

The study focused on three specific machine learning models: random survival forests, eXtreme Gradient Boosting (XGBoost), and deep neural networks, specifically the DeepSurv architecture. These were compared against the traditional Cox proportional hazards model. The performance of all models was evaluated using two key metrics: the C-index and the Integrated Brier Score.

The evaluation was performed under different data-generating mechanisms. The researchers varied several critical parameters, including sample sizes, censoring proportions, and the addition of noise variables. Furthermore, the study examined performance in the presence of different types of model misspecification.

The results indicated that all models showed improved predictive performance as sample sizes increased. However, the performance of all models declined as the proportion of censoring increased.

A significant finding was that an increase in noise variables reduced prediction accuracy across all models, regardless of the dataset size. In scenarios involving model misspecification and a large number of noise variables, tree-based models, such as random survival forests and XGBoost, demonstrated promising predictive performance compared to the Cox proportional hazards model and DeepSurv.

Conversely, the Cox proportional hazards model performed well when sample sizes were larger and the number of noise variables was fewer. It also maintained strong performance when the model was correctly specified or had only minor misspecification.

The authors conclude that the appropriate model for a specific medical application should be motivated by the research question and the underlying data structure.

Subject of Research: Comparative simulation study of Cox proportional hazards model and machine learning methods for time-to-event prediction

Article Title: A comprehensive simulation study evaluating the predictive performance of Cox proportional hazards model and machine learning methods for time-to-event data

Article References: Otieno, J. A., Häggström, J., & Eriksson, M. (2026). A comprehensive simulation study evaluating the predictive performance of Cox proportional hazards model and machine learning methods for time-to-event data. BMC Medical Research Methodology, 26(1), Article 240. https://doi.org/10.1186/s12874-026-03050-5

Image Credits: AI Generated

DOI: 10.1186/s12874-026-03050-5

Keywords: survival analysis, machine learning, Cox proportional hazards, simulation study, random survival forests, XGBoost, DeepSurv, time-to-event data, predictive performance, BMC Medical Research Methodology, comprehensive, simulation