Stochastic Explicit Calibration Algorithm for Survival Models

Calibration is essential for risk evaluation in various fields; including medicine, finance, and reliability analysis. Although extensive research has focused on calibration in classification and regression tasks using deep neural networks, survival analysis remains relatively underexplored, resulting in the lack of improved calibration methods. In particular, while previous work has proposed a calibration method for survival analysis, it relies on fixed bins, which can lead to biased calibration assessments and substantial loss of predictive accuracy in pursuit of calibration. This gap can hinder an accurate assessment of survival functions, leading to increased risk management costs. In this study, we introduce Stochastic Explicit Calibration (S-cal), an algorithm that employs random intervals instead of fixed bins, thereby advancing the calibration methods used in deep networks. The calibration performance of S-cal is evaluated using metrics specifically designed for handling censored data, such as D-calibration and the Kolmogorov-Smirnov metric. Extensive experiments on synthetic and real-world datasets demonstrate that S-cal consistently outperforms existing methods in terms of calibration accuracy. In addition, we highlight how improved calibration can improve downstream tasks, including optimizing resource allocation and improving patient care decisions. This work presents a significant advancement in the study of calibration for survival analysis, offering valuable information for more reliable risk assessment models.