Primary Submission Category: Machine Learning and Causal Inference
Dynamic Conformal Prediction of Survival with Time-varying Covariates
Authors: Yuyao Wang, Larry Han,
Presenting Author: Yuyao Wang*
Time-to-event prediction plays a central role in many causal inference problems, where stakeholders care not only whether an event will occur but also how long remains until it does. While conformal prediction provides distribution-free uncertainty quantification for time-to-event outcomes, existing methods often yield only one-sided lower prediction bounds and do not leverage time-varying covariates, limiting their usefulness for dynamic planning and decision-making. In this work, we develop two-sided dynamic conformal prediction intervals for individual event times among survivors that adapt to evolving covariate histories as new information becomes available. The proposed method achieves asymptotically valid marginal coverage at each decision time under the conditional independent censoring assumption given covariate history. Through simulation studies, we show that the resulting dynamic prediction intervals are substantially narrower on average than existing conformal survival approaches while maintaining nominal coverage. We further illustrate the method using publicly available benchmark time-to-event data sets.