A censoring-aware target interface for tabular foundation models in survival prediction

2026-07-13 · Source: stat.ML updates on arXiv.org · Field: Health & Wellbeing — Artificial Intelligence & Machine Learning, Data Science & Analytics, Health & Medical Research · Depth: Expert, extended

Summary

The SurvFM-RMST framework addresses right-censored time-to-event prediction in tabular patient data, a challenge for standard tabular foundation models that assume fully observed outcomes. This framework converts survival outcomes into jackknife pseudo-observation targets for Restricted Mean Survival Time (RMST) at a prespecified horizon. This approach enables multiple tabular backbones, including SurvFM-TabPFN, SurvFM-TabICL, SurvFM-TabDPT, SurvFM-TabH2O, and SurvFM-MITRA, to perform horizon-specific RMST regression without requiring survival-specific fine-tuning. Controlled simulations across 24 settings demonstrated accurate recovery of known conditional RMST, with pseudo-RMST targets outperforming naive alternatives. Benchmarking on 36 eligible static SurvSet datasets showed SurvFM-RMST backbones were competitive with established survival and RMST-regression comparators. The framework also supports retrospective patient stratification based on predicted RMST.

Key takeaway

For Machine Learning Engineers developing prognostic models with censored patient data, you should consider implementing the SurvFM-RMST framework. This approach enables you to utilize existing tabular foundation models by converting survival outcomes into pseudo-RMST targets, avoiding complex survival-specific fine-tuning. Validate your chosen backbone against specific data characteristics, runtime, and deployment needs, as performance varies. Use predicted RMST for robust patient stratification and event enrichment.

Key insights

SurvFM-RMST uses pseudo-observations to enable tabular foundation models for censored survival prediction.

Principles

• Censoring-aware target construction is crucial for survival prediction.
• Pseudo-RMST targets improve accuracy over naive approaches.
• Model performance varies by endpoint, horizon, and data constraints.

Method

SurvFM-RMST constructs jackknife pseudo-observation targets for RMST from censored data, then pairs these with tabular backbones for horizon-specific regression.

In practice

• Apply pseudo-RMST targets for censored time-to-event data.
• Validate backbone choice based on data, runtime, and deployment.
• Use predicted RMST for patient stratification and event enrichment.

Topics

• Tabular Foundation Models
• Survival Prediction
• Restricted Mean Survival Time
• Jackknife Pseudo-observations
• Censored Data
• Biomedical Decision Support

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation — Adapting tabular foundation models with a survival-aware head significantly improves clinical time-to-event prediction.
• A Practical Upper Bound on Selection Bias Effects in Medical Prediction Models — A novel upper bound estimates worst-case model performance under selection bias with partially observed data.
• CT-CLIP Representations for Multimodal Lung Cancer Survival Prediction — Domain-specific foundation models like CT-CLIP can enhance lung cancer survival prediction in data-scarce clinical environments.
• Semantic-Anchored Evidential Fusion for Domain-Robust Whole-Slide Survival Analysis — High-level pathology semantics provide domain-invariant representations crucial for robust whole-slide image analysis in cancer prognosis.
• Censoring-adjusted tree-based policy learning for estimating dynamic treatment regimes with censored outcomes — CA-TReL robustly learns dynamic treatment regimes from censored observational data using tree-based reinforcement learning.
• Proper Scoring Rules for Right-Censored Survival Data — A new framework enables proper scoring rules for right-censored survival data by mapping predictions through censoring mechanisms.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.