acceptodds
Withdrawn from ICLR 2027

Prototyping Progressive Survival-Risk into Cross-Modal Representation Learning for Multimodal Cancer Survival Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal cancer survival prediction integrates complementary information from pathological Whole Slide Images and other modalities, such as gene expression data, to estimate patient-specific survival risk. Existing methods mainly focus on individual-level cross-modal alignment, overlooking the shared risk structures among patients and the progressive nature of survival risk. To address these limitations, we propose Risk-Pro, a Progressive Survival-Risk Prototyping framework that introduces population-level risk structures into multimodal representation learning. Specifically, Risk-Pro constructs risk prototypes to capture shared prognostic patterns across patients with similar survival states and aligns patient representations with corresponding risk structures. Moreover, a Progressive Prototype Constraint is proposed to organize risk prototypes along a continuous survival trajectory, encouraging closer representations for adjacent risk states and greater separation for distant ones. Experiments on six TCGA cancer datasets demonstrate that Risk-Pro consistently outperforms state-of-the-art methods and effectively captures progressive survival-risk structures, as validated by risk trend evaluation, Kaplan-Meier survival analysis, and attribution analysis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Cancer Survival Prediction, Prototype Learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_bzU8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Risk-Pro for multimodal cancer survival prediction. It augments a pathology-genomics conditional VAE with risk-associated prototype supervision. Prototypes are constructed from average genomic features within outcome-defined patient groups, using only the training fold. The model reconstructs both individual genomic features and the corresponding group prototypes, while a risk-distance-weighted cosine objective regularizes relationships between prototypes. Cross-attention and gated pooling produce survival predictions. Experiments use five-fold cross-validation across six TCGA cancer cohorts, including complete-modality and gene-missing inference.

Critical strengths

The experimental evaluation covers six cancer cohorts and includes both complete-modality and gene-missing settings. The component ablations separately examine prototype reconstruction and the progressive constraint, and the paper also evaluates sensitivity to prototype granularity and loss weights.

Critical weaknesses
  1. The methodological advance over closely related risk-prototype methods is insufficiently established. The paper motivates its contribution by suggesting that existing approaches overlook population-level risk structures and that prior prototypes mainly compress features. However, PIBD (Zhang et al., ICLR 2024) already learns prototypes associated with risk bands defined by survival time and censoring status. The cross-modal generative backbone also follows LD-CVAE. The specific additions therefore appear to be outcome-group mean reconstruction targets and a risk-distance-weighted prototype regularizer. These differences need to be positioned more precisely. Comparing complete architectures in Table 1 does not isolate their contribution relative to established risk-prototype supervision applied within the same backbone.

  2. The central progressive-risk mechanism is not sufficiently justified. When minimized, Equation (8) penalizes cosine similarity for every pair of distinct prototypes, with larger weights for more distant risk levels. It contains no explicit adjacent-prototype attraction or relative-similarity ordering constraint. The authors needs to establish that its benefit comes specifically from correct risk ordering, rather than generic prototype separation.

Questions for authors
  1. What is the specific methodological distinction between the proposed risk-associated prototypes and PIBD’s risk-band-conditioned prototypes? Please add a matched-backbone comparison between the proposed prototype supervision and PIBD-style risk-prototype supervision to isolate the contribution of the proposed design.

  2. Please explain how minimizing Equation (8) produces the claimed ordinal organization. Can you compare the proposed distance weights against (i) equal weights for all prototype pairs and (ii) weights obtained after randomly permuting the risk-level order?

  3. How are censored patients with shorter follow-up handled in prototype construction and the prototype-reconstruction loss? What exactly defines “half of the overall time,” and what justifies the resulting “safe” designation? Please examine whether the improvements persist under an alternative grouping scheme that explicitly accounts for the incomplete outcomes of censored patients.

Reviewer_MSTFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Risk-Pro, a progressive survival-risk prototyping framework for multimodal cancer survival prediction using histopathological whole-slide images and genomic data. The method constructs population-level risk prototypes from genomic features of patients with similar survival outcomes and incorporates them into cross-modal representation learning through prototype-guided reconstruction. A Progressive Prototype Constraint is further introduced to organize risk prototypes according to their ordinal relationships. The framework also employs a conditional variational autoencoder to transfer genomic information to pathology representations, enabling pathology-only inference. Experiments on six TCGA cancer cohorts demonstrate competitive C-index performance. The authors additionally provide component ablations, risk-trend analyses, Kaplan–Meier survival curves, and attention-based visualizations to support the proposed approach.

Critical strengths

1. Clinically meaningful motivation. The paper addresses a relevant limitation of existing multimodal survival prediction methods by considering prognostic patterns shared across patients rather than focusing exclusively on patient-level cross-modal alignment. Incorporating population-level risk structures into representation learning is an interesting direction, particularly for learning from heterogeneous cancer cohorts.

2. Broad empirical evaluation. The experiments cover six TCGA cancer datasets and include comparisons with representative unimodal and multimodal survival prediction methods. The component ablations, risk-trend evaluations, Kaplan–Meier analyses, and visualization results provide complementary evidence supporting the potential utility of risk-aware prototype learning.

Critical weaknesses

1. Insufficient justification of the progressive risk formulation. The Progressive Prototype Constraint in Eq. (8) minimizes a distance-weighted cosine similarity between risk prototypes. While this encourages separation between distant risk levels, it does not explicitly encourage adjacent risk levels to remain close or guarantee monotonic similarity ordering. More importantly, the risk-level partition in Appendix A.7 relies on observed survival times and assigns certain censored patients to the safe group, potentially introducing inaccurate risk supervision. These issues directly affect the validity of the proposed population-level risk representation.

2. Unclear inference mechanism for missing genomic modalities. A central claim is that Risk-Pro supports pathology-only inference through cross-modal knowledge transfer. However, the C-VAE encoder in Eq. (19) explicitly requires genomic features as input, while the paper does not sufficiently specify how the latent genomic representation is obtained when those features are unavailable. This ambiguity makes it difficult to assess the reproducibility and validity of the pathology-only inference results.

Questions for authors

1. Progressive Prototype Constraint (Sec. 4.2). Eq. (8) minimizes distance-weighted cosine similarities between risk prototypes. How does this objective explicitly preserve higher similarity between adjacent risk levels while enforcing greater separation between distant ones? Could the authors provide a theoretical explanation and comparisons with ordinal ranking or triplet-based constraints?

2. Censoring-aware risk partition (Appendix A.7). The proposed risk grouping uses observed survival times and assigns censored patients with sufficiently long follow-up to the safe group. How are the remaining censored patients handled? Since censoring does not determine the actual survival outcome, could the authors investigate the sensitivity of risk prototypes to censoring rates and alternative censoring-aware partition strategies?

3. Pathology-only inference (Sec. 4.1, Appendix A.6). Eq. (19) defines the C-VAE encoder using both genomic and pathology features. How is the latent genomic representation inferred when genomic inputs are unavailable? Is a pathology-conditioned prior or an additional inference network employed? Please clarify the complete inference procedure and its implementation.

4. Statistical significance of performance improvements (Table 1). Although Risk-Pro achieves the highest reported C-index across six datasets, some margins are modest. For example, the improvement over LD-CVAE on UCEC is only 0.005 (0.712 versus 0.707), smaller than the reported standard deviations. Could the authors provide paired statistical tests or confidence intervals and clarify how the average improvement of 2.93% is calculated?

5. Biological interpretation of attention maps (Sec. 5.6). The authors interpret cross-attention patterns as evidence of associations between specific genomic pathways and pathological structures. However, attention weights alone do not establish pathway activity or biological mechanisms. Could these interpretations be supported by quantitative gene-expression correlations, pathway-level enrichment, or independent molecular validation?

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

Risk-Pro augments pathology–genomics survival prediction with outcome-grouped genomic prototypes, prototype reconstruction, and a distance-weighted prototype loss intended to encode an ordered survival-risk structure.

Critical strengths

The paper addresses a relevant problem and reports higher mean C-indices across all six TCGA cohorts using matched features and cross-validation splits. Component ablations, survival-stratification analyses, and experiments with genomic inputs removed provide useful supporting evidence.

Critical weaknesses

Equation (8) penalises similarity between all prototype pairs, including adjacent risk levels, without explicitly ensuring the claimed ordering or closeness of neighbouring states. The risk partition labels some censored patients as “safe” while leaving early-censored cases unspecified, raising concerns about the validity of the prototype supervision. Pathology-only inference is also insufficiently specified, since the stated encoder requires genomic inputs and the prior is not conditioned on pathology. My recommendation is Weak Reject because the promising empirical results do not resolve central concerns about the risk-ordering objective, censoring-aware supervision, and reproducibility of missing-modality inference.

Questions for authors

How are early-censored patients handled, and precisely how are genomic latent representations generated when genomic inputs are unavailable? Can the authors compare their progressive constraint against ordinary prototype separation and an explicit ordinal loss, and provide paired statistical evidence for the reported performance gains?

Reviewer_TDvAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Risk-Pro, built on LD-CVAE, for multimodal (WSI + gene expression) cancer survival prediction with optional pathology-only inference. It adds (i) risk prototypes, i.e. per-risk-level mean gene features that the decoder must also reconstruct, and (ii) a Progressive Prototype Constraint, an ordinal-distance-weighted cosine loss meant to order prototypes by risk level. On six TCGA cohorts under 5-fold CV it reports the best C-index, supported by ablations, a Jonckheere-Terpstra trend test, latent correlation heatmaps, Kaplan-Meier analysis and attention visualizations.

Critical strengths

The method is designed so that genomic data are used only during training and the model can still run with pathology only. The paper reports this gene-missing setting explicitly (Table 2, Fig. 4). This matters because genomic profiling is costly and often unavailable in practice, so a model whose pathology-only performance is reported is more useful than one that is evaluated only with complete data.

On six TCGA cohorts, Risk-Pro attains the highest mean C-index against unimodal, multimodal and LD-CVAE baselines, with the stated use of identical features and splits and a leakage precaution (prototypes built only from each training fold). The gains are modest relative to fold variance (see weaknesses), but the direction is consistent across cohorts, which is the main evidence that the proposed risk-level supervision may help.

Critical weaknesses

Eq. 8 is a weighted sum of cosine similarities with no positive term or margin, so it pushes all prototype pairs apart, including adjacent levels, rather than bringing adjacent levels closer as claimed. It also acts on encoded gene-feature means (through a linear encoder), not on the learned latents, and no analysis links it to the latent geometry. The supporting analyses do not isolate its effect: the Jonckheere-Terpstra test uses risk levels defined from the same survival outcomes the model is trained on, so any predictive model would show a trend, and no baseline (LD-CVAE, Intra-risk) is given. The correlation heatmaps (Fig. 3) are shown only for the full model, and several cohorts (BLCA, LUAD, UCEC, GBMLGG) do not show the described monotone trend. Since the "progressive" claim is the paper's main novelty, this undermines the contribution.

The gain over the best baseline is about 2.2 C-index points on average (0.005 on UCEC), small relative to fold standard deviations of 0.02-0.07, with no significance tests. The Progressive Constraint contributes only about 1.5 points in the ablation, within one standard deviation in every cohort, and on UCEC the prototype reconstruction alone does not help. K and the loss weights appear to be selected on the same CV folds used for reporting (no validation split is described). The closest prior work (ProSurv, MICCAI 2025, which also uses risk-interval prototypes and prototype-guided generation of missing genomics) is neither cited nor compared, nor are other missing-genomics methods cited in the introduction. The "Single-patient" ablation equals LD-CVAE, so the true contribution is the sum of the two additions, and a simple auxiliary risk-level (ordinal) supervision baseline is missing.

Questions for authors
  1. How exactly is every training patient assigned to a risk level (Appendix A.7)? In particular, what happens to censored patients with follow-up shorter than half of the "overall time", and what are the group sizes per fold?
  2. In Eq. 6, are the encoded prototypes and the encoded patient features used as reconstruction targets detached from the trainable gene encoder Ω_m? If not, how is a shrinkage solution avoided?
  3. How is z_gene obtained at test time when genes are missing (prior p(z|z_path) versus posterior mean)? Which patients (train, test, or pooled out-of-fold) are used for Figs. 2-4?
  4. How were K, λ1 and λ2 selected, and was the same tuning budget given to the baselines? Is there a validation split separate from the reported folds?
  5. How is the correlation between risk levels in Fig. 3 computed, and what is "the 2.93%" average improvement (reference method, absolute or relative)?

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

Related papers

Discussion (0)

Sign in to comment.