acceptodds
Under review as a conference paper at ICLR 2027

From Partial Likelihood to Pairwise Preference: Deep Survival Analysis with Bradley–Terry Models

Abstract

Deep learning has become central to modern survival analysis, with applications spanning clinical trials, operations research, and economics. A defining—and largely underexplored—feature of many such applications, most acutely in medical imaging, is that the covariate dimension far exceeds the number of available subjects (): a single CT volume already contains on the order of voxels, whereas a cohort rarely exceeds a few thousand patients. This regime forces very small training batches, which interact poorly with the Cox proportional hazards model: its partial likelihood couples all subjects through a shared risk-set denominator, making minibatch gradients biased and stochastic optimization unreliable exactly when batches are smallest. We make a simple observation with broad consequences: fitting a Bradley–Terry model in place of the Cox partial likelihood recasts survival training as a sum over independent pairwise preferences, yielding unbiased minibatch gradients while targeting the same risk ordering. Casting survival analysis as preference learning further lets us import the extensive toolbox developed around the Bradley–Terry model in RLHF—graded (fine-grained) labels that exploit the magnitude of survival-time gaps. On a 3D medical imaging dataset, our method matches or surpasses Cox-based deep survival baselines, with the largest gains in the small-batch, high-dimensional regime where partial-likelihood training degrades most.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Survival Analysis, Deep Survival Analysis, Preference Optimization

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

This paper studies how to train models that predict which patients are at higher risk of an event, such as death, when each 3D medical image is so large that only a few patients fit in a training batch. Instead of using the usual Cox loss, the paper trains the model by comparing pairs of patients whose event order is known. It explains when these comparisons are valid despite incomplete follow-up and shows that small batches can estimate the gradient of the new pairwise loss without bias. The paper also explores giving stronger labels to patient pairs with larger observed time gaps.

Critical strengths
  1. The problem is practical and worth studying. Training on full-size CT images makes ordinary Cox training difficult because each patient's loss involves other patients who may not fit in the same batch.

  2. The pairwise method is easy to understand, and the paper clearly explains how to use comparisons even when some patients' final event times are unknown. I also appreciate the experiment on full-resolution 3D CT images.

Critical weaknesses
  1. I am not convinced that the main method is sufficiently different from earlier work. Cox-CC and BigSurvSGD already allow stochastic training using closely related pairwise comparisons. The paper needs to show what is gained by its particular way of selecting patients and combining their pairs.

  2. The experiments do not yet establish a clear advantage. The CT results are competitive but mixed across batch sizes. I also find the claim about stronger labels based on time gaps unconvincing: a large observed gap does not necessarily mean a large difference in patients' risks, and the evaluation does not clearly show improved survival-time predictions.

Questions for authors
  1. Section 2 and Algorithm 1: Unless I am missing something, BigSurvSGD already uses a pairwise survival loss and stochastic training, while Cox-CC can also train with minibatches. What exactly is new here? Can you show that comparing all available pairs within a batch works better than these existing ways of sampling pairs? The description of Cox-CC as needing all events at every update also seems inaccurate.

  2. Table 1: The paper says that when a batch contains only two patients, minibatched Cox and Bradley-Terry use the same pairwise loss. Why, then, are their parameter-recovery errors different? Is this because the losses are scaled differently or because of other training settings?

  3. Section 5.2: Why should a longer observed time gap mean a stronger difference in risk? Two patients could have similar relative risks but very different absolute survival times, and one patient may stop being followed before the event happens. Could the improvement simply come from using softer labels, rather than from the actual time gaps?

  4. Section 3.3 and Appendix D, Table 4: The soft labels lower the reported error in predicted survival time, but do not improve the integrated Brier score, which evaluates survival predictions over time. Also, the time-error calculation uses only patients whose event was observed. How would the conclusion change if the evaluation also accounted for patients whose follow-up ended before the event?

3 comments

We thank the reviewer for highlighting the practical relevance of the problem and the simplicity of our approach. We have updated the manuscript in response to your feedback, clarifying the relationship to existing methods, adding baseline comparisons, and revising the evaluation and interpretation of graded preference labels.

Across the evaluated batch sizes, BT at b=4b=4 achieves the highest mean Harrell’s C, Uno’s C, and IPCW-AUC among the reported configurations (Table 2). At this batch size, it also yields the largest estimated high-versus-low hazard ratios across the two-, three-, and four-group stratification analyses (Appendix E, Figures 2–4). Below, we address your four questions.

  1. Relationship to Cox-CC and BigSurvSGD. We explicitly acknowledge the shared pairwise logistic loss when Cox-CC uses one control or BigSurvSGD uses strata of size two. Our practical distinction concerns uniform patient sampling and the reuse of encoded volumes across all comparable pairs:

    “Building on these foundations, we study a training procedure that samples patients uniformly and uses all comparable pairs within each minibatch. Each volume is encoded once per update, and its scalar risk score is reused across comparisons.”

    We now provide direct comparisons in both the parameter-recovery and whole-volume CT experiments. On the high-dimensional DBCD and CHOP cohorts, BT achieves lower mean relative ℓ2\ell_2 error than Cox-CC and BigSurvSGD for b≥4b\geq4 (Table 1). For example, on DBCD at b=16b=16, the errors are 1.809 for BT, 2.709 for BigSurvSGD, and 13.331 for Cox-CC. The imaging comparisons use the same backbone and hardware, with the strongest mean discrimination results obtained by BT at b=4b=4 (Table 2) and comparable training times across these methods (Table 5).

  2. Different parameter-recovery results at b=2b=2. Thank you for highlighting this detail. At b=2b=2, the two methods share the same pairwise loss, but their implementations use different normalization constants, as shown in our code. This changes the gradient scaling and can lead to different fitted estimates.

  3. Motivation for graded labels and the role of label smoothing. We clarified in Appendix D that graded labels are a heuristic: nearby observed times receive softer targets, while larger gaps receive stronger targets. As the revised manuscript states:

    “We therefore do not interpret an observed time gap as a direct measure of the latent risk difference or claim that our graded targets follow from the proportional-hazards assumption.”

    When the later observation is censored, the observed gap is only a lower bound on the true event-time separation. We also agree that softer labels themselves may contribute to the improvement. The present experiments evaluate the combined graded-label procedure; they do not isolate the contribution of time-dependent grading from that of label smoothing.

  4. Censoring-aware evaluation and interpretation of IBS. We added IPCW weighting to the RMST prediction-error evaluation and defined it explicitly in Section 3.3, Eq. (4). Under this evaluation, soft BT retains lower mean RMST error than hard BT on all three cohorts (Table 4): improvements of 1.34 months on METABRIC, 6.6 days on SUPPORT, and 1.34 months on GBSG. Mean Harrell’s and Uno’s concordance differ by at most 0.002.

    IBS is slightly higher under soft BT, and SUPPORT IPCW-AUC decreases. We therefore revised our conclusion accordingly:

    “The results therefore support improved accuracy under the IPCW-weighted time-prediction metric, while largely preserving concordance; they do not establish a uniform improvement in survival-probability calibration.”

We hope these revisions clarify the practical contribution and provide a more precise interpretation of the empirical results.

The response clarifies the b = 2 results and adds useful comparisons with Cox-CC and BigSurvSGD. However, the authors agree that the loss is the same as in prior work, and the CT gains appear only at one batch size. They also admit that the graded labels are a heuristic and do not separate their effect from label smoothing, while IBS gets slightly worse. The main concerns therefore remain, and therefore I maintain my original score.

Thank you for your follow-up. We would like to respectfully clarify that we did not state or agree that our loss is the same as those used in prior work. The fact that both objectives contain pairwise logistic terms does not make the overall losses identical. This distinction is fundamental to our formulation and is explicitly discussed in the manuscript.

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

This paper proposes replacing the Cox partial likelihood with a Bradley–Terry (BT) pairwise loss for deep survival analysis. The authors establish the connection between pairwise survival ordering and the BT model under proportional hazards, and show that the proposed objective admits unbiased minibatch gradients. They also explore graded preference labels based on observed survival-time differences.

Critical strengths

The paper addresses a practical optimization problem in deep survival analysis, particularly for high-dimensional medical imaging where batch sizes are limited. The proposed objective is simple to implement, and the theoretical derivation under censoring is clearly presented.

Critical weaknesses

My main concern is the limited methodological novelty. The proposed pairwise logistic loss is essentially the same as the one-control nested case-control formulation used in Cox-CC (Kvamme et al., 2019), which has already been developed for scalable deep survival training. The main difference appears to be how comparable pairs are sampled and weighted. I do not see a substantial methodological contribution beyond this change.

I also find the motivation unconvincing. The authors repeatedly emphasize that minibatch Cox gradients are biased relative to the full partial likelihood, but this does not imply inconsistent estimation. Zeng et al. (2026, JASA) have already established consistency and convergence results for minibatch Cox estimation. The unbiasedness of the BT gradient with respect to its own objective does not establish an advantage over minibatch Cox.

Finally, the empirical improvements are modest, particularly on the main CT dataset, where discrimination is largely comparable across methods. The graded-label extension also lacks a clear statistical justification, and its reported improvements are difficult to interpret given the evaluation metric used.

Questions for authors
  1. Relation to existing nested case-control methods: How does the proposed method differ from Cox-CC (Kvamme et al., 2019) beyond the sampling and weighting of comparable pairs? Since both use the same pairwise logistic loss, what is the benefit of the proposed sampling strategy? I would expect a direct comparison with Cox-CC on the main imaging dataset under matched computational budgets.

  2. Motivation and theoretical contribution: How do the authors reconcile their motivation with the consistency results for minibatch Cox estimation established by Zeng et al. (2026)? Since Proposition 2 only establishes unbiased gradients for a different objective, what does it tell us about the statistical properties of the resulting estimator? A comparison of the estimators, rather than their gradient bias relative to different objectives, would be more informative.

  3. Graded labels and empirical evaluation: What justifies using observed time differences to determine preference strength under proportional hazards, especially when the later observation is censored? Also, RMST-RMSE evaluated only on uncensored subjects is not a reliable measure of survival calibration. Can the authors provide a censoring-aware evaluation of the claimed improvement? The IBS results in Table 4 do not show an improvement from graded labels.

2 comments

We thank the reviewer for recognizing the practical utility of our approach. We have updated the manuscript in response to your feedback, clarifying its relationship to prior work, adding direct baseline comparisons, and revising the evaluation and interpretation of graded preference labels.

  1. Relationship to Cox-CC and direct comparisons. We expanded Section 2 to acknowledge that Cox-CC with one control shares the pairwise logistic loss. Our practical contribution concerns patient sampling and the reuse of encoded volumes across all comparable within-batch pairs. The revised manuscript states:

    “Building on these foundations, we study a training procedure that samples patients uniformly and uses all comparable pairs within each minibatch. Each volume is encoded once per update, and its scalar risk score is reused across comparisons.”

    We also added Cox-CC and BigSurvSGD to the whole-volume CT experiments using the same 3D CNN backbone and hardware (Table 2). At b=4b=4, BT achieves higher mean Harrell’s C than Cox-CC (0.629 versus 0.593) and higher IPCW-AUC (0.661 versus 0.609). The updated stratification analyses include both baselines, with BT yielding the largest estimated high-versus-low hazard ratios across the two-, three-, and four-group evaluations (Appendix E, Figures 2–4). Table 5 shows comparable training times for BT and Cox-CC.

  2. Relationship to Zeng et al. and comparison of estimators. We appreciate this reference and have incorporated it into Section 2. We agree that gradient bias relative to the full partial likelihood does not imply inconsistent estimation. As the revised manuscript explains:

    “Their analysis permits small fixed minibatches, but its stated asymptotic framework considers increasing sample size with a fixed input dimension, so that d/n→0d/n\to0.”

    Our work addresses complementary questions about finite-sample training with high-dimensional volumetric inputs and limited cohort sizes. Its primary contribution is applied: developing a simple deep-learning training procedure for whole-volume 3D survival prediction. The imaging studies added to Sections 1 and 2 provide further context for the practical memory constraints that motivate patch-based processing or very small patient minibatches.

    We therefore compare the fitted estimators directly through relative ℓ2\ell_2 error (Table 1), cosine similarity (Table 3), and SGD training trajectories (Figure 1). On the high-dimensional DBCD and CHOP cohorts, BT achieves lower mean relative ℓ2\ell_2 error than minibatch Cox for b≥4b\geq4, while attaining comparable cosine similarity with the ground-truth parameter. These experiments examine finite-sample estimation behavior, rather than inferring an advantage solely from gradient unbiasedness.

  3. Graded-label motivation and censoring-aware evaluation. We clarified in Appendix D that graded labels are a heuristic, motivated by assigning softer supervision to pairs with nearby observed times. We do not claim that observed time gaps determine latent risk differences: patients with similar risks can have substantially different event times. When the later observation is censored, its observed gap is only a lower bound on the true event-time separation. The IPCW adjustment concerns evaluation and does not correct this limitation of the training labels.

    We added the IPCW-weighted RMST prediction-error definition to Section 3.3, Eq. (4), and updated Table 4 accordingly. Soft BT reduces mean error relative to hard BT on all three cohorts: by 1.34 months on METABRIC, 6.6 days on SUPPORT, and 1.34 months on GBSG. Changes in mean Harrell’s and Uno’s concordance are at most 0.002. However, IBS is slightly higher on all three cohorts, and SUPPORT IPCW-AUC decreases from 0.597 to 0.582. We have therefore revised the interpretation explicitly:

    “The results therefore support improved accuracy under the IPCW-weighted time-prediction metric, while largely preserving concordance; they do not establish a uniform improvement in survival-probability calibration.”

We hope these revisions clarify the practical contribution, its empirical support, and the scope of our conclusions.

Thank you for the additional experiments and clarifications. However, my main concern about the novelty still remains. The proposed method essentially uses the same pairwise loss as Cox-CC, and the main difference seems to be the sampling strategy and the use of all comparable pairs within each batch. Although the additional experiments show some improvements, I am still not convinced that this represents a substantial methodological contribution beyond existing methods. Therefore, I will maintain my original score.

Reviewer_uNP2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work proposes training deep survival models with a Bradley-Terry (BT) loss over comparable patient pairs instead of the Cox partial likelihood. Under proportional hazards and conditionally independent censoring, it derives a logistic pairwise ordering probability and shows that uniformly sampled patient minibatches give unbiased gradients of the pairwise objective. It also explores soft preference labels based on observed survival-time gaps. Experiments include semi-synthetic parameter recovery, survival prediction from 3D CT, and graded-label comparisons on three tabular cohorts.

Critical strengths
  1. The formulation is simple and clear with details provided, making proposed training procedure easy to understand and implement.
  2. The imaging experiments also address a relevant memory constraint with useful controls. Training directly on large CT volumes makes small batches a practical concern. Using the same 3D backbone and GroupNorm across losses helps separate changes in the objective from batch-normalization effects. Reporting multiple data splits and several discrimination metrics is also useful.
Critical weaknesses
  1. I am concerned that the advance over existing pairwise survival methods is unclear. BigSurvSGD already provides a pairwise logistic objective, its U-statistic formulation, and parameter equivalence under proportional hazards (the result left as a conjecture in Section 4.1). Any benefit from using all within-batch pairs therefore needs direct evaluation.

  2. Currently the imaging results do not establish a consistent advantage. DeepHit has higher mean concordance at (b=8), while Cox has lower RMST-RMSE at (b=4). The stratification analysis also averages scores from different fitted models without explaining how their arbitrary additive offsets were aligned. Since patients receive predictions from different subsets of models, this could distort the pooled ranking.

Questions for authors

Please refer to the weakness section for my concerns.

1 comment

We appreciate the reviewer’s recognition of the practical memory constraints that motivate our experimental design. Your feedback helped us update our manuscript, clarify the relationship to prior work, strengthen the empirical comparisons, and explain our evaluation procedure.

  1. Relationship to Cox-CC and BigSurvSGD. We expanded Section 2 to acknowledge the shared pairwise logistic loss and clarify the differences in sampling and reuse of encoded patients. Cox-CC samples controls from each selected event patient’s risk set, while BigSurvSGD optimizes partial-likelihood objectives over random strata. Our procedure uses all comparable pairs within a uniformly sampled patient minibatch. As the revised manuscript states:

    “Building on these foundations, we study a training procedure that samples patients uniformly and uses all comparable pairs within each minibatch. Each volume is encoded once per update, and its scalar risk score is reused across comparisons.”

    This makes effective use of the few volumes that fit in memory, allowing each expensive encoding to contribute to multiple comparisons.

  2. Direct comparisons on whole-volume CT data. We added Cox-CC and BigSurvSGD to the imaging evaluation using the same 3D CNN backbone (Section 5.1, Table 2). At a global batch size of four, BT achieves the highest mean Harrell’s and Uno’s C-indices and IPCW-AUC among all evaluated configurations. BT also improves on Cox-minibatch and BigSurvSGD in all reported mean discrimination metrics at both batch sizes under similar computational budget.

  3. Alignment of pooled risk scores. Our aggregation procedure already rank-normalized each fitted model’s held-out scores before averaging; we have now made this explicit in Section 5.1 and Appendix E. Each patient’s normalized scores are averaged only across splits in which that patient was held out, and each patient contributes once to the KM analysis. Consequently, arbitrary model-specific offsets and positive rescaling do not affect the aggregated scores, even when patients receive predictions from different subsets of models. The updated stratification plots also include Cox-CC and BigSurvSGD. At b=4b=4, BT yields the largest estimated high-versus-low hazard ratios in the two-, three-, and four-group analyses (Figures 2–4).

We acknowledge that performance depends on the metric and batch size. At b=8b=8, DeepHit has higher mean concordance, while BT has lower RMST-RMSE (932 versus 1018 days). Cox also has slightly lower RMST-RMSE at b=4b=4 (959 versus 964 days). Our findings support a simple and effective training procedure, with the best reported mean discrimination achieved by BT at b=4b=4 and the largest estimated high-versus-low hazard ratios in the reported risk-group analyses.

Comments on the paper

1 comment

We thank the reviewers for their initial feedback. We hope our responses have clarified their questions, and we look forward to hearing their thoughts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.