Censoring-Aware Pseudo-Labeling for Semi-Supervised Survival Prediction using Risk-Order Agreement Filtering

Abstract

Practical cohorts often have a small labelled fraction with confirmed time-to-event outcomes and a larger covariate pool with missing, delayed, or incomplete endpoints; yet, survival prediction is essential to clinical prognosis modeling. A problem formulation is derived that makes censoring constraints and pseudo-label safety requirements explicit after reviewing survival modeling, assessment under censoring, and semi-supervised learning (SSL) ideas that are relevant to pseudolabeling. Here is a conservative method blueprint called Adaptive Pseudo-Label Selection and Correction (APSC). It uses mediancrossing pseudo events and pseudo-censoring at tmax to generate weak pseudo outcomes from predicted survival curves. A curvegeometry confidence proxy is used to admit pseudo labels. Unstable pseudo labels are filtered using risk-order agree-ment across rounds. Finally, a down-weighted Cox partial-likelihood objective is used to inject pseudo supervision. Where naive selftraining can breach censoring limits and magnify early model bias, the formulation aims at registry-style situations with delayed endpoints and heterogeneous follow-up. In order to provide a supervised backup in the event that pseudo labels are unreliable, the survey condenses these factors into specific design criteria and deployment diagnostics. There are no numerical comparisons reported in the text; instead, the focus is on auditoriented operating methods like validation gating, monitoring of acceptance and correction rates, and tests for coefficient stability