Conformal prediction with censored data using Kaplan-Meier method

In this paper, we introduce a prediction algorithm founded on conformal prediction, tailored for constructing prediction intervals in the context of censored survival data. Conformal prediction frameworks distinguish themselves from other prediction paradigms by their non-empirical evaluation, reliance on user-defined confidence intervals for modeling errors, and widespread adoption across regression and classification methodologies, inclusive of survival analysis, in recent years. Herein, we present a novel application wherein the Kaplan-Meier method is employed to compute empirical quantiles of nonconformal scores, specifically tailored for censored schematic variables. This novel approach facilitates the generation of well-calibrated prediction intervals for survival times, augmenting any existing survival prediction algorithm. Validation of its efficacy and computational efficiency is performed on both the real-world dataset 'SUPPORT' and the synthetic dataset 'RRNLNPH.'