CURE-OOD: Benchmarking Out-of-Distribution Detection for Survival Prediction

Wenjie Zhao, Jia Li, Mingrui Liu, Jing Wang, Yunhui GuoarXiv 2026 · arXiv

TL;DR

This paper introduces CURE-OOD, the first benchmark designed to systematically evaluate out-of-distribution (OOD) detection in cancer survival prediction using real-world clinical CT data. The authors reveal that standard classification-based OOD detectors fail in survival contexts due to inverted logit dynamics and propose HazardDev, a simple, survival-native baseline that leverages predicted hazard curves to reliably flag shifted samples.

What Problem Does This Paper Address?

Deep learning models have shown great promise in predicting patient survival outcomes from medical images like computed tomography (CT) scans. However, in real-world clinical workflows, CT scans are collected over long periods, across different hospitals, and using various scanner models or protocols. This variability introduces covariate shifts—systematic changes in image appearance (such as resolution, noise, and contrast) while the underlying clinical prediction task remains fixed.

Figure 1: CT data distribution shifts caused by cross- institution variability and scanner changes over time.

While OOD detection is heavily studied in standard image classification, survival prediction is fundamentally different: its outputs are time-indexed risk or survival distributions rather than simple class confidences. Until now, no benchmark existed to study how these acquisition-induced distribution shifts affect survival models, or whether conventional OOD detectors can reliably identify them.

Methods

The authors evaluate survival models using Multi-Task Logistic Regression (MTLR), a popular framework that reformulates survival analysis into a sequence of dependent binary classification subtasks across discrete time intervals.

To address the unique challenges of survival modeling, they introduce HazardDev, a simple survival-aware reference baseline for OOD detection. Instead of relying on raw logits or class confidences, HazardDev operates directly on a survival-native quantity: the predicted hazard curve (how event risk evolves over time). During training, the model establishes an expected average hazard profile. At test time, HazardDev assigns an OOD score by calculating how much a new sample's predicted hazard curve deviates from this baseline training profile.

Datasets

The benchmark is constructed using the RADCURE cohort, which contains radiotherapy planning CT scans from 2,366 head and neck cancer (HNC) patients collected between 2005 and 2017.

CURE-OOD partitions this dataset into non-overlapping training, in-distribution (ID) test, and OOD test splits based on four key CT acquisition parameters found in DICOM metadata:

  1. Pixel spacing (spatial sampling resolution)
  2. Exposure time (photon accumulation)
  3. Slice thickness (through-plane averaging)
  4. X-ray tube current (intensity)
Figure 3: Examples showing how acquisition parameters affect CT image appearance. Larger pixel spacing covers a wider field of view with decreased resolution, longer ex- posure time and higher tube current reduce noise, and thicker slice thickness produces smoother but blurrier im- ages.

The benchmark evaluates performance across four distinct clinical survival endpoints: overall survival (OS), local failure-free survival (LFFS), regional failure-free survival (RFFS), and distant failure-free survival (DFFS).

Key Insights and Ideas

  • Classification OOD Detectors Fail (and Invert): Mainstream logit-based OOD detectors (such as Energy, ASH, and SCALE) perform poorly in survival prediction. In many settings, they exhibit reversed ordering, mistakenly assigning more OOD-like scores to perfectly normal ID samples.
  • The "Why" Behind the Failure: In standard classification, ID samples yield high-magnitude, confident logits. However, in survival MTLR models, the training objective forces ID logits to be strongly negative at early intervals (reflecting high confidence that the patient survives). When encountering an unfamiliar OOD sample, the model's uncertainty causes these logits to drift upward toward zero. Because standard detectors assume larger/closer-to-zero logit magnitudes imply ID confidence, they suffer from systematic ID-as-OOD errors.
  • Hazard Curves as Robust Signals: Across all evaluated shifts, OOD samples consistently exhibit systematically higher predicted hazard curves than ID samples. Operating on these survival-native predictive quantities provides a much more stable OOD signal than transferring classification confidence scores.

Key Results

  • Downstream Performance Degradation: Covariate shifts noticeably harm survival prediction accuracy. When ID-trained models were evaluated on OOD data, the concordance index (C-index) dropped across tasks, with the most severe drop reaching 0.2987 in distant failure-free survival (DFFS) under exposure time shifts.
  • HazardDev Superiority: HazardDev consistently outperformed classification-style detectors across different visual backbones (ViT and ResNet). On the overall survival (OS) task using a ViT backbone, HazardDev achieved an average AUROC of 0.5873, outperforming mainstream classification detectors which often scored below 0.50 (worse than random guessing due to the inverted logit problem).

Limitations and Future Work

The authors point out a critical limitation: OOD detectability does not perfectly correlate with downstream performance risk. Some distribution shifts are easy for detectors to isolate but only modestly reduce survival prediction accuracy, while others severely degrade accuracy but remain difficult to separate from ID data. The authors suggest that future research should move beyond generic anomaly detection and focus on developing task-aligned OOD methods that specifically flag shifts detrimental to clinical decision-making.

Summary updated

Abstract

``How long can I live and remain free of cancer?'' is often the first question a patient asks after receiving a cancer diagnosis and treatment. Accurate survival prediction helps alleviate psychological distress and supports risk stratification and personalized treatment planning. Recent survival prediction frameworks have shown strong performance using computed tomography (CT) images. However, variations in imaging acquisition introduce out-of-distribution (OOD) samples caused by covariate shifts that undermine model reliability. Despite this challenge, to our knowledge, no existing benchmark systematically studies OOD detection in cancer survival prediction. To address this gap, we introduce the Cancer sURvival bEnchmark for OOD Detection (CURE-OOD), the first benchmark for systematically evaluating OOD detection in survival prediction under controlled acquisition-induced distribution shifts. CURE-OOD defines scanner-parameter-based training, in-distribution (ID), and OOD test splits across four survival prediction tasks. Our experiments show that covariate shifts notably reduce survival prediction performance. It also shows that mainstream classification-oriented OOD detectors can fail in survival prediction. Finally, we include HazardDev as a simple survival-aware reference baseline for OOD detection. CURE-OOD enables systematic analysis of how distribution shifts affect both downstream survival performance and OOD detectability.
BibTeX
@misc{zhao2026cure,
  author = {Wenjie Zhao and Jia Li and Mingrui Liu and Jing Wang and Yunhui Guo},
  title = {{CURE-OOD: Benchmarking Out-of-Distribution Detection for Survival Prediction}},
  year = {2026},
  eprint = {2605.00350},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.00350},
}
Figure