Simpler Calibration for Survival Analysis
Survival analysis, also known as time-to-event analysis, is the problem to predict the distribution of the time of the occurrence of an event. This problem has applications in various fields such as healthcare, security, and finance. While there have been many neural network models proposed for survival analysis, none of them are calibrated. This means that the average of the predicted distribution is different from the actual distribution in the dataset. Therefore, X-CAL has recently been proposed for the calibration, which is supposed to be used as a regularization term in the loss function of a neural network. X-CAL is formulated on the basis of the widely used definition of calibration for distribution regression. In this work, we propose new calibration definitions for distribution regression and survival analysis, and demonstrate a simpler alternative to X-CAL based on the new calibration definition for survival analysis.
PDF Abstract
