New arXiv framework benchmarks competing risks survival models with SHAP extension

Modeling survival with competing risks—where one event like cancer death prevents another, such as cardiovascular death—has grown increasingly complex. A new arXiv submission from Begoña B. Sierra and colleagues (including Peter S. Hall and Catalina A. Vallejos) addresses the lack of standardization in this space. The authors built a reproducible, extensible benchmarking framework that allows survival models to be compared head-to-head across multiple datasets. Instead of relying on a single metric, the framework evaluates four key aspects: calibration, discrimination, overall prediction error, and clinical utility. That breadth provides a more complete picture of real-world model performance than typical accuracy-only comparisons.

A notable addition is an extension of SHAP (SHapley Additive exPlanations) designed specifically for competing risks. This gives practitioners a model-agnostic way to interpret covariates' contributions over time, showing how risk factors influence different event types as time progresses. All code is open-source on GitHub, and the paper includes 23 pages of main text plus 31 pages of supplementary information, with 7 figures illustrating the framework's application. For clinicians and researchers in oncology, cardiology, and epidemiology, this framework lowers the barrier to rigorous model evaluation and could become a standard reference point for future survival analysis methodologies.