Research map / Learning theory and algorithms / Prediction and decision losses

History-Aware Conformal Prediction Sets for Censored Time-to-Event Outcomes

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Conformal prediction sets · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HAPS, a split-conformal framework that builds two-sided prediction intervals for event times among survivors at a decision time using time-varying covariate histories, with IPCW censoring adjustment, plus doubly robust variants (HAPS-DR, HAPS-A).

Problem

Conformal prediction for individual event times under right censoring is static (baseline covariates only), yielding wide, uninformative intervals for individuals who have survived to a decision time; censoring depends on time-varying covariates, which existing methods do not adjust for.

Approach

Calibrate a scalar parameter indexing nested quantile-based interval candidates at the decision time using IPCW (and AIPCW/post-processing doubly robust) estimating equations built on time-varying covariate histories, targeting survivor-conditional coverage.

Key results

• In simulations, HAPS reduces median interval length by 34-75% relative to baselines at tau=6 while achieving near-nominal (~90%) coverage; at tau=3 reductions are 5-23%.
• On colon and PBC data at year 5, HAPS and HAPS-DR reduce median interval length by 13-60% (colon) and 9-22% (PBC) relative to baseline methods.
• HAPS-DR tends to overcover in simulations, while HAPS undercovers under misspecified censoring models.

Surprising findings

• Baseline IPCW and trunc overcover at later prediction times, and even at tau=0 baselines show coverage gaps because baseline-only censoring adjustment is insufficient.
• The baseline-only static intervals are close to zero for lower bounds even with oracle quantiles, showing baseline covariates alone are too uninformative.

Limitations

• Validity relies on conditional independent censoring given observed covariate history and positivity; violations may give invalid coverage.
• HAPS undercovers under censoring model misspecification; HAPS-DR is conservative, especially with small samples or poorly trained prediction models.
• HAPS-A requires estimating additional nuisance functions (h_tau), which is computationally demanding and can be unstable in small samples.

Experimental setup

Datasets: colon (survival R package) pbcseq (survival R package) simulated DGM1 simulated DGM2

Benchmarks: simulation DGM1 and DGM2 colon pbcseq

Metrics: empirical survivor-conditional coverage median prediction interval length IPCW estimate of survivor-conditional coverage

Baselines: IPCW trunc LM

Builds on: conformal prediction inverse probability of censoring weighting Dynamic-DeepHit AIPCW conformalized survival analysis

Closest papers by meaning

• Toward Conditional Distribution Calibration in Survival Prediction (NeurIPS 2024 · similarity 0.87)
• Doubly Robust Conformalized Survival Analysis with Right-Censored Data (ICML 2025 · similarity 0.86)
• Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.86)
• Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction (NeurIPS 2026 · similarity 0.85)
• Conformal Prediction for Causal Effects of Continuous Treatments (NeurIPS 2025 · similarity 0.85)
• Conformalized Survival Distributions: A Generic Post-Process to Increase Calibration (ICML 2024 · similarity 0.85)