A large-scale neutral comparison study of survival models on low-dimensional data

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Established benchmarks for survival data are limited in scale and scope.
• Existing studies often lack rigorous tuning and evaluation procedures.
• This study addresses the need for a neutral, large-scale benchmark for survival data analysis.

Purpose of the Study:

• To neutrally evaluate a wide range of survival models.
• To provide generalizable guidelines for practitioners using survival data.
• To establish a comprehensive benchmark for single-event, right-censored, low-dimensional survival data.

Main Methods:

• Benchmarked 21 models (classical statistics to machine learning) on 34 public datasets.
• Tuned models using Harrell's C-index and Integrated Survival Brier Score.
• Evaluated models across six metrics: discrimination, calibration, and overall predictive performance.

Main Results:

• No single method statistically significantly outperformed the Cox proportional hazards model for tuning measures.
• Oblique random survival forests and likelihood-based boosting showed superior average ranks in overall predictive performance.
• Boosting, tree-based methods, and parametric survival models ranked higher in discrimination.

Conclusions:

• The Cox proportional hazards model is a reliable default for low-dimensional, right-censored survival data.
• More flexible machine learning methods may be advantageous depending on dataset characteristics.
• This benchmark provides a foundation for future survival model research and application.