A unified imputation framework for interval-censored data: comparing AFT, RSF, and DeepSurv models [comunicação oral]

Abstract

Interval-censored data are common in longitudinal studies and pose challenges for time-to-event analysis. This work proposes a unified imputation-based framework for handling interval-censored data, where latent event times are iteratively generated within the observed censoring intervals and the censoring mechanism is handled externally through a scaled redistribution procedure. Within this framework, different predictive models—including AFT, Random Survival Forests, and DeepSurv—can be consistently compared through an iterative imputation scheme based on pseudo-event times within the observed intervals, followed by a common scaled redistribution procedure. Performance is assessed through simulations under varying censoring levels, interval widths, and hazard distributions, with extensions to nonlinear effects and high-dimensional covariates. Results are further validated using real-world clinical datasets.