SPARC: SURVIVAL PSEUDO-LABEL ADAPTIVE RE- FINEMENT AND CALIBRATION

ICLR 2026 Conference Submission24861 Authors

Keywords: Survival Analysis, Pseudo-Label Refinement, Calibration, Censoring, Long-Term Survival Prediction, Deep Ensemble Learning, Weibull–Kaplan–Meier Initialization, Clinical Reliability, Oncology, Healthcare AI

TL;DR: We propose an adaptive pseudo-label refinement and calibration framework that enables accurate, biologically consistent long-term survival prediction despite heavy censoring and limited follow-up.

Abstract: Accurate survival prediction is critical for oncology, public health, and reliability engineering, yet existing methods remain constrained by limited follow-up, heavy censoring, and static pseudo-labeling practices. In many clinical datasets, including Our reconstructed cohort of $N = 50{,}155$ patients with observed follow-up of only 74.742 months (58.8\% deceased, 41.2\% censored), long-term outcomes remain unobserved, preventing reliable 10-year (120-month) survival estimation. We address this gap by introducing a dynamic pseudo-label refinement and calibration framework that transforms incomplete follow-up into extended, biologically consistent survival trajectories. Starting from a hybrid Weibull–Kaplan–Meier initialization, pseudo-labels are iteratively corrected under survival-theoretic constraints and clinical plausibility rules, including enforcing zero survival beyond death and monotonic survival probabilities for censored patients. These refined labels are propagated through a deep ensemble trained with variance-penalizing objectives and monitored via diagnostic feedback for stability and uncertainty calibration. This process enables survival labels to evolve adaptively, rather than remain static preprocessing artifacts, and produces clinically plausible estimates well beyond the observed horizon. We applied to the 50{,}155-patient cohort, the framework achieved rapid convergence and outstanding predictive performance ($R^2 = 0.9964$, MAE = 0.0066, C-index = 0.9915), with predictions tightly calibrated, biologically consistent, and robust under long-term censoring. We validated Our proposed framework on two public datasets of N = 2509 \& N = 205 available in \cite{Metabric_Kaggle} (Metabric) \& \cite{finalfit2023survival} (Malignant Melanoma) achieved remarkable results ($R^2 = 0.9924$ \& $0.9781$, MAE = 0.0142 \& 0.0247, C-index = 0.9633 \& 0.8459) by follow-up to 480 \& 240 months respectively. Thus, by bridging the 74.742-month follow-up limit with reliable 120-month projections on Our dataset, Our work establishes adaptive pseudo-label refinement as a principled foundation for long-horizon, interpretable, and clinically reliable survival modeling. Moreover, we are going to publicly publish Our dataset and code at \url{https://doi.org/10.5281/zenodo.17163267} and \url{https://anonymous.4open.science/r/Dynamic-Pseudo-Labeling-D2AB/} respectively for the research community.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 24861