Conformalized Survival Distributions: A Generic Post-Process to Increase Calibration

The paper presents a novel Conformalized Survival Distributions (CSD) method for improving the calibration of survival analysis models without sacrificing discrimination. CSD, a post-processing technique, provides theoretical guarantees and demonstrates its effectiveness across 11 real-world datasets, showcasing its practical value and robustness. It adapts conformal regression to censored data, is model-agnostic, and can be applied to various survival models. The study emphasizes the importance of maintaining discrimination while enhancing reliability of survival probability predictions by aligning them with actual event distributions. CSD separates calibration and discrimination during optimization, focusing on discrimination first and then refining predictions. It improves calibration metrics like D-cal and KM-cal while preserving the model's ranking ability (C-index). The method is compared to existing techniques, showing consistent improvements in calibration while maintaining or enhancing discrimination. The paper also addresses different calibration measures, such as D-calibration, 1-cal, and KM-calibration, and discusses the trade-offs between discrimination and calibration in survival analysis.