acceptodds
Under review as a conference paper at ICLR 2027

SurvBridge: Survival Representation Transfer Across Heterogeneous Clinical Cohorts

Abstract

Heterogeneous clinical cohorts often differ in available variables, observation processes, and follow-up, complicating the transfer of survival representations learned from large electronic health records (EHRs). We introduce SurvBridge, a schema-flexible framework for cross-cohort survival representation transfer that combines feature-semantic encoding, survival-aware multi-source contrastive pre-training, explicit representation of missingness state, and Cox-based survival adaptation. SurvBridge supports pre-training across multiple EHR sources with and without time-to-event labels, without requiring identical feature spaces, and adapts the learned representation to target-native clinical schemas. We evaluate SurvBridge on three public clinical survival datasets and two independent cardiovascular cohorts. SurvBridge achieves strong survival discrimination against classical, deep survival, and tabular foundation-model baselines. Pre-training provides its largest gains when labeled target data are scarce, with SurvVPCL providing more consistent low-resource transfer than the alternative survival-aware contrastive objectives evaluated. On naturally incomplete clinical data, explicitly representing missingness state improves mean discrimination over deterministic missing-value handling and changes how partially observed clinical variables contribute to predicted risk. These findings position cross-cohort survival pre-training primarily as a means of improving sample efficiency while accommodating heterogeneous clinical feature spaces and missingness representation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
survival analysis, transfer learning, electronic health records, cardiovascular mortality, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SfnMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SurvBridge transfers survival representations across clinical cohorts with partially overlapping feature schemas. It encodes feature identity with a frozen BioLORD encoder, adds learned missingness embeddings, pretrains with multi-source contrastive learning on MIMIC-IV, eICU, HiRID and an unlabeled Kaggle cohort using a survival-weighted negative term, then fine-tunes a Cox head. It is evaluated on two private cohorts (URRAH, IHD) and three public ones (SUPPORT2, SMARTO, FLCHAIN) using Ctd, plus low-resource and convergence analyses.

Critical strengths

The paper is well engineered and the evaluation is broad by this subfield's standards: the baseline suite includes tabular foundation models, the low-resource curve and convergence-time table are informative, and the paired-seed analyses are more careful than usual. Ethics documentation for the PhysioNet data is explicit.

Critical weaknesses
  • The headline gains sit within seed noise: on URRAH, SurvBridge (Scratch) scores 0.855 [0.849, 0.861] against SurvTRACE at 0.851 [0.843, 0.858], a difference of +0.004, and on IHD, VPCL scores 0.803 against Scratch at 0.800.
  • The missingness contribution is +0.003 to +0.006 with seed confidence intervals of +/-0.004 to 0.016; all tests are seed-level (n = 5), and no patient-level confidence intervals or tests are reported, so the repeated-seed design cannot separate method effects from run variance.
  • The abstract's "strong survival discrimination" does not hold on SUPPORT2, where Scratch (0.755) is below CoxPH (0.769) and TabPFN-MTLR (0.774).
  • The URRAH missingness rates differ between Fig. 4 and Fig. 7 for the same features (AF 89.2% vs 86.7%, BUN 53.0% vs 47.7%, Heart Rate 13.7% vs 17.2%), and for Alcohol the rank order flips (20.0% vs 27.9%).
  • In Table 5, the "BERT" row equals Table 1's SurvBridge (Scratch) exactly and "BioLORD" equals SurvBridge (SurvVPCL), so the encoder ablation is confounded with the pretraining objective.
  • The variance reported in Fig. 8 does not match its own axes (text 55.7% for PC1+PC2 versus 36.2% + 22.2% = 58.4% plotted), and Fig. 10/11 time axes for URRAH and IHD imply a 23x difference while the reported medians differ by only 1.8x, suggesting mixed units.
  • The baseline comparison is unequal on the missingness dimension: baselines use median imputation while SurvBridge uses learned missingness embeddings, and URRAH is 39.6% missing, so the comparison partially measures imputation strategy rather than transfer.
  • TransTab-Cox appears in the results but is absent from the hyperparameter search spaces in Table 3.
Questions for authors
  1. Why do URRAH missingness rates differ between Fig. 4 and Fig. 7, and which is correct?
  2. Does Table 5 hold pretraining fixed when swapping encoders, or are those rows mislabeled?
  3. Can you report patient-level confidence intervals and paired tests rather than seed-level ones?
  4. What are the units in Fig. 10/11, and is the URRAH axis correct?
  5. Are MIMIC-IV-Surv and eICU-Surv patients disjoint from the pretraining pools?
1 comment

Thank you very much for your feedback. We include below answers to your questions. We have uploaded a new version of the paper (and a pdf with changes in blue as part of the supplementary material).

Q1: Figure 4 reports the missingness rate on the first seed’s training split, while Figure 7 contains the entire dataset.

Q2: Unfortunately there was a mixup between results of the final models and a previous version. We have replaced the second row in table 5 with the correct model version. Now, only the first row in table 5 is equal to the SurvBridge (Scratch) row in table 1, as should have been from the beginning. Thank you for noticing the issue.

Q3: Unfortunately due to time constraints we were not able to recompute patient level CIs. However, seed level intervals also incorporate some patient-level variability, since for different seeds the test datasets change.

Q4: The figures used months for the URRAH dataset and days for IHD. We changed the figure to always report days.

Q5: Yes, the MIMIC-IV and eICU patients used for downstream survival prediction are not used for pretraining.

General comments: We have rephrased the text in general to clarify better that the superiority of our method is not for all datasets, and that SurvVPCL is not better than VPCL. We underline that SurvVPCL was not the only novelty of our model: SurvBridge also incorporates missing value embeddings and combines TransTab-style encoding with the Cox survival prediction head, demonstrating overall superiority compared to the state of the art in almost all datasets studied, even without pre-training (scratch). Indeed, missingness representation is an important improvement over the state of the art. Regarding figure 8, there was an error reporting the explained variance, we have fixed it. Adjusted the number to 58.4% Regarding TransTab-Cox, we have added the details to Table 3, we mistakenly excluded it.

Reviewer_HKGtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

They propose SurvBridge, which enables transfer learning for survival representations across clinical cohorts with different schemas. They use a frozen text encoder (BioLORD) for feature names embeddings , and missing values are represented with a learned embedding. The encoder used here is pre-trained with SurvVPCL, a survival-aware variant of vertical-partition contrastive learning, and then adapted to the target with a Cox head. They use MIMIC-IV, eICU and HiRID for pertaining, plus an unlabeled cardiovascular dataset (KaggleCD). They also evaluate it on two cardiovascular cohorts (URRAH, IHD) and three survival analysis datasets (SUPPORT2, SMARTO, FLCHAIN) with classical baselines and deep learning based models like PFN-based survival models. They claim SurvBridge is at least as good as the baselines and pertaining mostly improves sample efficiency.

Critical strengths

Building a general survival model from large public EHRs with different schemas is an interesting problem for the field, since differences in follow-up scale and missingness patterns are a real challenge in practice. Also the low-resource results in table 6 are the most convincing finding and yet burred in the appendix (though it lacks classical baselines as also mentioned below), I mean showing SurvVPCL has an strong performance on small-data baselines is very useful practical message and should be emphasized more.

Critical weaknesses

1- The results in table 5 vs table 1 shows that the BERT row is exactly the same as SurvBridge Scratch, and the BioLORD row is exactly the same as SurvBridge (SurvVPCL), which is weird and needs discussion. So, the reported delta therefore equals the gain from pre-training, not from changing the encoder, which contradicts the conclusion, so “consistently improves” is not supported or is overclaimed here.

2- The novelty is a bit thin here SurvVPCL is closely resembles the VPCL and also the results show no reliable differences between them. I mean, as I understand here equation 6 changes a weight only when both patients in a pair had an observed event which should be pretty rare for the the pertaining datasets.

3- The result in figure 3 are within noise distance of a trivial alternative model and I cant see a strong baseline for that. So as I understand, the learned embedding beats zero-constant filling by only 0.001–0.006? an standard baseline is missing here and therefore “consistently improves vs deterministic imputation” doesnt mean the learned embedding is better than simple indicator approaches.

Questions for authors

1- In Table 5, why are the BERT and BioLORD rows identical to the Scratch and SurvVPCL rows of Table 1?

2- Can you report the Table1 baselines with a explicit missing-indicator features?

3- Can you report CoxPH (penalized), RSF, and the rest on the same nested subsets at N=100–2000?

1 comment

Thank you very much for your feedback. We include below answers to your questions. We have uploaded a new version of the paper (and a pdf with changes in blue as part of the supplementary material).

Q1: Thank you for catching this one. Unfortunately there was a mixup between results of the final models and a previous version. We have replaced the second row in table 5 with the correct model version. Now, only the first row in table 5 is equal to the SurvBridge (Scratch) row in table 1, as should have been from the beginning.

Q2: We have added CoxPH to the missingness analysis in Figure 3. In the low resource setting we note that CoxPH is better than scratch, and comparable to the pre-trained SurvBridge version, especially for constant zero filling, which appears to be the best deterministic strategy. In the large resource setting SurvBridge becomes superior. Unfortunately due to time constraints, we were not able to study all other baselines.

Q3: We have added CoxPH with zero filling to the low-resource analysis in Figure 2 and Table 6. Interestingly, the simpler model appears to provide good performance in the low-resource setting, close to the pre-trained SurvBridge (even if still lower on URRAH, probably due to missingness embedding). On URRAH, as the training dataset size increases, the performance becomes worse than our Scratch. Unfortunately, due to time constraints we were not able to study all other baselines.

General comments: we have rephrased the text in general to clarify better that the superiority of our method is not for all datasets, and that SurvVPCL is not better than VPCL. We underline that in SurvVPCL, weight wijw_ij is lower than 1 only when pairs have a similar time to event and come from the same dataset, because batches are drawn from one dataset at a time. Indeed, this means downweighting is quite rare, possibly explaining the low difference between the methods. We also underline that SurvVPCL was not the only novelty of our model: SurvBridge also incorporates missing value embeddings and combines TransTab-style encoding with the Cox survival prediction head, demonstrating overall superiority compared to the state of the art in almost all datasets studied.

Reviewer_JXGeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SurvBridge, a framework for survival representation transfer across clinical cohorts with heterogeneous feature spaces and missingness patterns. It combines feature-semantic encoding, VPCL-based contrastive pre-training with survival-informed negative weighting (SurvVPCL), and Cox-based fine-tuning on target cohorts. Experiments on five survival datasets show that pre-training improves predictive performance, particularly when target training data are limited.

Critical strengths

The paper addresses a relevant problem, as large EHR datasets could be useful for survival prediction in smaller clinical cohorts. The experiments cover multiple source and target datasets, and the low-resource results show substantial benefits from pre-training compared with training from scratch.

Critical weaknesses

My main concern is the motivation for the proposed SurvVPCL objective. The method simply downweights negative pairs based on the proximity of observed event-time quantiles. However, observed event-time proximity does not necessarily imply similarity in underlying survival risk. It is unclear why reducing contrastive repulsion between such patients should lead to better representations for downstream relative-risk prediction. The weighting also ignores censored patients, even though their observed follow-up provides partial information about event times.

More importantly, the experiments provide little evidence that this survival-specific modification is useful. Standard VPCL already achieves nearly identical performance, including in the low-resource settings where pre-training is most beneficial. Most of the reported gains therefore appear to come from the existing contrastive pre-training framework rather than the proposed SurvVPCL objective.

The paper also emphasizes heterogeneity across clinical cohorts, but does not sufficiently address how differences in patient populations, survival endpoints, and censoring processes affect transfer. Sharing an encoder across sources does not necessarily make the learned survival information transferable, especially when the underlying prognostic relationships may differ.

Questions for authors
  1. What is the statistical motivation for using observed event-time quantiles to weight contrastive negatives? How does this weighting relate to similarity in underlying survival risk, and why is the information from censored observations excluded? The authors should provide a more convincing justification for this objective.

  2. Since VPCL and SurvVPCL perform almost identically, can the authors better isolate the contribution of survival supervision? In particular, how much of the contrastive objective is actually affected by the weighting, given the low event rates in the source cohorts? An experiment with shuffled event times would help determine whether the observed effect comes from meaningful survival information.

  3. How are the survival endpoints, time origins, and censoring rules defined in MIMIC-IV, eICU, and HiRID? Given that the cohorts may involve different outcomes and populations, what justifies transferring their survival supervision to the target tasks?

  4. Can the authors analyze the contribution of individual source cohorts, for example through leave-one-source-out experiments? This would help establish whether heterogeneous multi-source training is consistently beneficial or whether some sources introduce negative transfer.

1 comment

Thank you very much for your feedback. We include the answers to your questions. We have uploaded a new version of the paper (and a pdf with changes in blue as part of the supplementary material).

Q1: The motivation for including SurvVPCL, along with traditional VPCL, was to test empirically whether including some survival information in the pre-training phase could improve survival prediction downstream. Indeed, a similar time-to-event does not guarantee that patients are similar, but since we do aim to predict time-to-event later, we hoped that this change could allow for uncovering a survival-specific representation during pre-training. As you correctly underline, the improvement is not statistically significant and too limited (see response to Q2 below). We underline the fact that the weight wijw_{ij} is different from 1 only if patients have different event times within the same source dataset (because our pre-training batches are dataset-specific). This ensures that negative pairs are always from the same dataset, and therefore from the same time-to-event distribution. If patients come from different datasets or are censored, the weight becomes 1 as in standard VPCL. This means censored patients are not excluded but included as in standard VPCL. We have edited the methodology to underline these aspects.

Q2: Indeed SurvVPCL does not show significant improvement over VPCL. We have therefore rephrased our objectives, results and discussion to make it more clear that our framework, SurvBridge, improves over SOTA baselines mostly due to missingness representation and, in low resource settings, due to VPCL.

Q3: The distribution of follow-up times is shown in the supplementary material, section A2. Indeed survival events are different among datasets, motivating the fact that SurvVPCL weights negatives only within the same source datasets. In any case, we do observe good transfer in the low resource setting. This is due to the fact that VPCL builds a generic clinical representation that turns out to be useful for survival analysis.We have edited the paper to clarify these aspects.

Q4: We thank the reviewer for this comment. In the supplementary material we included an analysis of different pre-training dataset combinations (Table 8): large-scale EHR, SurvSet-Clinic and MIMIC-IV only, under the full resource setting. Small differences are visible when changing pre-training settings, and generally the advantage over training from scratch is maintained. Furthermore, pretraining speeds up target adaptation. Unfortunately due to time constraints we were not able to perform the full leave-one-out experiment. We included this analysis in the future work paragraph.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.