A0719
Title: Conformal prediction for time-to-event outcome subject to truncation
Authors: Weidong Wang - University of Massachusetts Amherst (United States)
Rebecca Betensky - New York University (United States)
Jing Qian - University of Massachusetts Amherst (United States) [presenting]
Abstract: Accurate predictions of clinically meaningful times-to-event are crucial for optimal resource allocation, such as determining insurance coverage for expensive Alzheimer's disease drugs and assessing eligibility for clinical trials. Conformal prediction offers a flexible framework for quantifying uncertainty using arbitrary prediction algorithms, yielding prediction intervals with valid marginal coverage without distributional assumptions. While conformal prediction methods for censored time-to-event data have been developed recently, they have not been developed for truncated data. This is addressed by developing a conformal prediction method for left-truncated and right-censored data, facilitated by inverse probability of censoring and truncation weighting. This approach is extended to accommodate other types of truncation, as well, such as right, double, and sequential truncation. Simulations and semi-synthetic examples demonstrate the effectiveness and robustness of our methods in different settings, supporting their usefulness in biomedical applications with truncated and censored data. An R package, cfTrunc, is developed to facilitate implementation.
