Distribution-Free Lower Predictive Bounds for Right-Censored Time-to-Event Data via Hybrid Quantile Learning and DFT-Adaptive Conformal Calibration

ICLR 2026 Conference Submission21303 Authors

Keywords: conformal prediction, censored data, random forest, kernel estimation, model-free prediction

Abstract: Reliable uncertainty quantification for time-to-event outcomes is challenging when observations are censored and censoring depends on covariates. While conformal prediction offers a distribution-free tool, existing methods for right-censored data typically rely on fixed, global filtering rules that ignore how censoring varies across individuals. We introduce a hybrid, model-agnostic framework that combines flexible conditional quantile learning with a Data-Filtered Threshold-adaptive (DFT-adaptive) conformal calibration scheme. A base learner, instantiated as censored quantile regression forests, is trained with censoring addressed via localized Kaplan-Meier estimation; and conformity scores are calibrated nonparametrically using covariate-dependent-censoring thresholds. Our development yields marginally valid lower predictive bound that adapts to heterogeneous censoring and scales to nonlinear settings without parametric assumptions on the censoring mechanism. We provide theoretical guarantees and supporting experiments to demonstrate that the method effectively delivers adaptive, interpretable, distribution-free uncertainty quantification for censored outcomes.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 21303