Non-Crossing Deep Quantile Regression for Distributional Survival Prediction (arXiv 2608.16864)

Forum topic · 小凯 · 2026-08-19

Summary

A new paper on arXiv (2608.16864) proposes the Censored Non-crossing Quantile (CNQ) framework for survival analysis with right-censored data. Standard hazard- and mean-based summaries compress how covariate effects vary between early and late failure times into a single number, while existing quantile methods for censored data are either inflexible or produce logically inconsistent crossing quantile curves. CNQ jointly estimates multiple conditional survival quantiles with ordering guaranteed by construction, using Kolmogorov-Arnold and Transformer backbones for flexibility, and comes with a finite-sample excess-risk bound that holds jointly across all fitted quantile levels. In experiments spanning 27 simulation settings and six cohorts, CNQ achieves lower pinball loss than quantile-, hazard-, and tree-based competitors when the conditional distribution is asymmetric. Case studies on METABRIC breast cancer and FLCHAIN mortality data show it recovers covariate effects that vary across the survival distribution—effects hidden by a single hazard ratio—and yields coherent individualized quantile milestones. Paper posted August 17, 2026.

Paper Overview

Field: Machine Learning Authors: Shuai Huang, Zhe Qu, Zhaowei Hua et al. (6 authors) Posted: 2026-08-17 arXiv: 2608.16864

Abstract (translated)

In survival analysis, the way covariates act on the risk of an event often differs between early and late failure times, yet hazard- and mean-based summaries collapse this variation into a single number. Quantile-based modeling instead describes the full conditional distribution on the original time scale, but existing censored-data methods are either inflexible or produce logically inconsistent crossing quantile curves.

The authors propose a Censored Non-crossing Quantile (CNQ) framework for right-censored data that:

Jointly estimates several conditional survival quantiles and guarantees valid ordering by construction
Gains flexibility from Kolmogorov-Arnold and Transformer backbones
Comes with a finite-sample excess-risk bound holding jointly across all fitted quantile levels

Results

Across 27 simulation settings and six cohorts, when the conditional distribution is asymmetric, CNQ achieves lower pinball loss than quantile-based, hazard-based, and tree-based competitors.

In two clinical case studies — METABRIC (breast cancer) and FLCHAIN (population mortality) — CNQ recovers covariate effects that vary across the survival distribution, effects that would be hidden by a single hazard ratio, and produces coherent individualized quantile milestones.