{"id":"05ab33f4-d4d7-444a-aee6-26d9fc83f083","arxiv_id":"2505.11997","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MRePath uses sheaf hypergraphs on pathology patches plus dynamic weighting to fuse genomics, reaching a mean C-Index of 71.5%, about 3 points above prior multimodal baselines on five TCGA datasets.","lead":"This paper introduces MRePath, a model that predicts cancer survival by combining whole-slide pathology images with genomic data using hypergraph learning and a modality-rebalancing mechanism. The authors report that it ranks patient risk more accurately than several prior methods on five public TCGA cancer datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3.0-3.4 point C-Index gain may be inflated because k and the architecture were selected on the same folds used for the reported results; a nested validation check is needed.","rationale":"The reader's weakest assumption points to the same load-bearing issue: the hyperedge threshold k is selected after seeing full benchmark results without nested cross-validation. My reading agrees, and I extend it slightly: the same concern applies to the choice of hyperedge types, fusion strategy, and encoder, all of which are ablated on the same folds used for the final reported numbers. This is a correctness-risk issue because the headline margin (3.0-3.4 points) is comparable to the per-fold variability, so the selection protocol could materially change the conclusion. The paper does provide code, extensive ablations, and a coherent method, which is real evidence, but these do not resolve the selection-bias concern. I would keep the reader's CONDITIONAL verdict rather than moving to REJECT or UNVERDICTED, because a nested validation run is a straightforward fix and the method itself is not fundamentally flawed. The proposed concrete test would either confirm the reported 71.5% or reveal that the margin is an artifact of test-set selection; until then, the quantitative claim should be treated as provisional.","tokens_in":24611,"tokens_out":4730,"duration_ms":52626,"concrete_test":"Re-run the k-ablation with nested selection: for each of the five outer folds, split the training 80% into inner train/validation, select k from {0,4,9,24,48} (and optionally the hyperedge/fusion choices) on the inner validation, then evaluate the selected model on the held-out outer fold. Report the mean test C-Index and the paired difference versus PIBD. If the nested mean falls below roughly 70% or the margin over PIBD shrinks to below 1 point, the 71.5% headline reflects selection bias rather than a genuine advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is Table 1's 71.5% mean C-Index and the 3.0-3.4 point margin over PIBD/MOTCat. The support for that claim is weakened by the model-selection protocol. Section 4.3, Figure 3, and Table 6 choose k=9 after inspecting per-dataset C-Index results from the 5-fold splits that are then reused to report the final numbers. The same holds for the architectural choices in Tables 2, 3, and 4 (hyperedge types, fusion strategy, encoder). Because no held-out data or nested selection is used, the reported mean is effectively a maximum over a set of configurations evaluated on the benchmark, not an unbiased estimate for a single pre-specified model. With per-fold standard deviations of 1.9-5.8 points and a 3.0 point margin over PIBD, this selection effect alone could explain the headline improvement. The Limitations section additionally concedes that a single k has different spatial scope across slides, so the choice is data-dependent rather than a scale-invariant default. No significance testing is reported, so even the reproduced baseline margins are not shown to be robust.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MRePath, a multimodal survival prediction framework that combines sheaf-hypergraph learning on pathology whole-slide images with a dynamic modality-rebalance mechanism for pathology-genomics fusion. The pathology branch constructs topological and feature-based hyperedges over patch-level features and applies a sheaf hypergraph convolution to capture contextual and hierarchical structure; the genomic branch embeds selected gene categories with an MLP. A dynamic weighting scheme computes mono- and holo-confidence scores to reweight the modalities, followed by an interactive alignment fusion with cross-attention. The method is evaluated on five TCGA cancer cohorts (BLCA, BRCA, CO-READ, HNSC, STAD) using 5-fold cross-validation, reporting a mean C-Index of 71.5%, which is 3.0-3.4 percentage points above the best compared baselines (PIBD and MOTCat). The paper also includes ablations on hyperedge types, the threshold k, fusion strategies, and pathology encoders, plus Kaplan-Meier stratification curves.","tokens_in":24830,"tokens_out":3044,"duration_ms":30877,"significance":"If the reported gains are robust, the paper makes a useful empirical contribution: it demonstrates that hypergraph-based patch aggregation and explicit modality reweighting can improve multimodal survival prediction over standard MIL and late-fusion baselines. The authors provide code, evaluate on five public datasets, include per-fold standard deviations, and perform ablations for each proposed module. The Kaplan-Meier analyses with log-rank p-values below 0.05 on all five datasets are also a positive sign of discriminative ability. However, the central claim hinges on whether the reported C-Index advantage is an unbiased estimate or an artifact of model selection on the same folds used for evaluation; this is the main load-bearing concern.","major_comments":[{"comment":"The hyperedge construction threshold k is selected after inspecting per-dataset C-Index results on the same 5-fold splits that are subsequently used to report the final numbers. The main text states that increasing k enhances performance with the best results at k=9, and Table 6 in the supplementary shows that k=9 is chosen across all datasets even though other values (e.g., k=48 for CO-READ, k=4 for STAD) sometimes achieve higher means with larger variance. Because no nested cross-validation or held-out selection is used, the reported 71.5% mean C-Index is effectively the maximum over a set of configurations evaluated on the test folds, not an unbiased estimate for a prespecified model. This is load-bearing because the margin over PIBD is only 3.0 points while per-fold standard deviations range from 1.9 to 5.8 points. The same concern applies to the pathology encoder selection in Table 4, where the five encoders produce mean C-Indices from 71.5% to 71.9%, a spread far smaller than the per-fold variability.","section":"Section 4.3, Figure 3, Table 6"},{"comment":"No significance testing is reported for the comparison with baselines. The C-Index tables report mean and standard deviation over five folds, but with only five folds and overlapping patient splits across methods, the 3.0-3.4 point advantage over PIBD/MOTCat may not be statistically reliable. The authors should report paired tests across folds (e.g., Wilcoxon signed-rank or permutation tests) or bootstrap confidence intervals for the per-dataset and mean C-Index differences. Without such tests, the claim that MRePath \"consistently outperformed\" all methods is not fully supported by the reported uncertainty.","section":"Section 4.2, Table 1"},{"comment":"The paper acknowledges that a fixed k gives different spatial scopes for slides with very different patch counts, which means the chosen k=9 is data-dependent rather than a scale-invariant default. This limitation interacts with the selection issue: the global threshold is tuned on the benchmark, yet its meaning varies across slides, so the reported generalization may not transfer to cohorts with substantially different WSI sizes or patch densities. The authors should either adapt k per slide based on a principled criterion or evaluate sensitivity across datasets with a proper nested protocol.","section":"Section 5, Limitations"}],"minor_comments":[{"comment":"The abstract claims \"over 3.4%\" improvement, but Table 1 shows 3.0% over PIBD and 3.4% over MOTCat. The claim should be stated more precisely, e.g., \"3.0-3.4 percentage points\".","section":"Abstract"},{"comment":"The holo-confidence formula uses logarithms of wm_p and wm_g without specifying constraints that these values are positive. If mono-confidence outputs are not guaranteed to be strictly positive (e.g., after a linear layer or ReLU), the log is undefined. The authors should clarify the activation and numerical safeguards.","section":"Section 3.3, Eq. (8)"},{"comment":"The text says Conch \"emerges as the most effective patch feature encoder,\" but Table 4 shows mean C-Indices of 71.8 (UNI), 71.9 (Conch), 71.7 (Phikon2), 71.5 (CTransPath), and 71.5 (ResNet50). These differences are negligible relative to the reported standard deviations; the claim should be softened or supported with significance analysis.","section":"Section 4.3, Table 4"},{"comment":"The threshold values are inconsistent across the paper: the main text and Figure 3 use k=0,5,9,25,49, while the supplementary Table 6 uses k=0,4,9,24,48. Please align the notation and reported grid.","section":"Supplementary, Table 6"},{"comment":"There is a typo: \"our resluts\" should be \"our results.\"","section":"Supplementary, Section 8.2"},{"comment":"The evaluation section says results are reported as mean C-Index ± STD \"across the five datasets,\" which is ambiguous: it should clarify that the mean is over the five folds within each dataset, and the overall mean is the average of the five per-dataset means.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical study, but the headline result is compromised by the model-selection protocol: the hyperedge threshold and encoder are tuned on the same folds used for the final evaluation, and no significance testing is provided. This is fixable with nested cross-validation or by reporting validation-based selection and then applying the chosen configuration to a truly held-out test split. The methodological contribution is incremental but the empirical study is careful in other respects. I would not reject; I would request a major revision with a properly unbiased evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a serious, well-structured incremental paper with released code, five TCGA datasets, and a thorough ablation set. Second, the headline C-Index margin (71.5 vs 68.5 for PIBD) is likely inflated, because the key hyperparameter k and the pathology encoder were chosen after looking at the same five folds that produce the final numbers. That is the dominant fact for reading it.\n\nWhat's actually new: it brings sheaf hypergraphs to WSI survival analysis, constructs both topological and feature hyperedges, and adds a dynamic mono/holo-confidence reweighting plus interactive co-attention fusion. Each piece has roots in earlier work (Di et al., Duta et al., Jaume et al.), but this combination, and the evaluation on BLCA, BRCA, CO-READ, HNSC, STAD, is new. The paper deserves credit for the ablation structure, which separates hypergraph type, threshold, fusion strategy, and encoder, and for explicitly acknowledging in the Limitations that a fixed k gives different spatial scope across slides.\n\nThe soft spots are real and load-bearing. In Section 4.3, Figure 3 and Table 6 pick k=9 after inspecting per-split C-Index values on the same splits reused for Table 1. The encoder choice in Table 4 is similarly post-hoc; several encoders beat ResNet50 on individual datasets, yet ResNet50 is used for the main results. With per-fold standard deviations of 1.9–5.8 points and a 3.0 point margin over PIBD, even a small selection effect could account for the whole gain. No significance testing is reported, and some baselines (SNN+CLAM, Porpoise) are cited rather than reproduced, so the comparison pipeline may not be fully matched. The abstract's '>3.4%' also doesn't match the mean 3.0% margin; minor, but should be cleaned.\n\nNone of this makes the method a dud. The mechanism is plausible, the ablations are internally consistent, and the authors are upfront about the k limitation. The paper would likely survive honest re-evaluation with a smaller, more believable gain.\n\nWho this is for: anyone working on multimodal computational pathology, hypergraph models of WSIs, or modality imbalance in fusion. It's a useful methodological reference and a good case study for why nested model selection matters in survival benchmarks.\n\nRecommendation: send it to peer review, but require a nested validation or a true held-out test set, significance testing or at least confidence intervals, and clearly matched or reproduced baselines. I'd want a stats-savvy referee on it.","headline":"Useful incremental method with honest ablations, but the headline C-Index gain is likely inflated by selecting k and the encoder on the same folds used for the final numbers.","tokens_in":25380,"tokens_out":3181,"would_cite":true,"duration_ms":29685,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MRePath claims 71.5% mean C-Index for cancer survival prediction by combining sheaf-hypergraph pathology features with dynamic modality rebalancing.","keywords":["survival prediction","whole slide images","hypergraph learning","sheaf hypergraph","modality imbalance","multimodal fusion","cancer prognosis","C-Index"],"falsifier":"Re-run the same 5-fold protocol with k selected per training fold, either nested or on a held-out validation split, and compare the mean C-Index against PIBD and MOTCat; if the 3-point gap shrinks to within one standard deviation, the central superiority claim is not supported.","tokens_in":24408,"feed_emoji":"🧬","tokens_out":5770,"duration_ms":53170,"temperature":0.7,"pith_summary":"This paper argues that survival prediction from whole-slide pathology images plus genomic profiles is held back by two correctable defects: multi-instance learning aggregates patches without preserving contextual and hierarchical structure, and pathology features numerically dominate genomics during fusion. To fix both, the authors propose MRePath, which models each slide as a hypergraph whose hyperedges connect patches both by spatial proximity and by feature similarity, then propagates information through a sheaf hypergraph Laplacian. A dynamic weighting module estimates each modality's reliability alone and in interaction, and an interactive cross-attention fusion rebalances the two modalities before risk prediction. On five TCGA cancer cohorts, MRePath reports a mean C-Index of 71.5%, about 3 percentage points above the strongest previous multimodal methods compared, PIBD and MOTCat. If correct, the result suggests that higher-order patch relationships and explicit modality rebalancing are worthwhile design choices for prognostic pathology-genomics models.","feed_headline":"Hypergraph modeling lifts cancer survival C-Index to 71.5%","feed_subtitle":"Pathology-genomic rebalance beats prior multimodal baselines by 3 points across five TCGA cohorts.","key_machinery":"The load-bearing machinery is the sheaf hypergraph on the pathology side and the mono/holo-confidence weighting on the fusion side. A hypergraph is a graph in which one hyperedge can connect many nodes at once; here nodes are tissue patches and hyperedges are formed either from spatial proximity in the topological space or from embedding similarity in the feature space, so each hyperedge encodes a local neighbourhood or a recurring morphological pattern. The sheaf hypergraph replaces the ordinary graph Laplacian with a sheaf Laplacian that assigns linear maps between nodes and hyperedges, letting information flow with structure rather than simple averaging, which is what the paper credits for preserving contextual and hierarchical detail. On the fusion side, mono-confidence scores are per-modality reliability estimates produced by MLPs, and holo-confidence is a ratio of log mono-confidences that quantifies how complementary the modalities are; a softmax over their combination produces the final modality weights, and residual co-attention layers then align the reweighted features. The dynamic weighting is presented as a plug-and-play adjustment that could be attached to other multimodal survival architectures.","core_discovery":"The central claim is that replacing patch aggregation with hypergraph learning and counteracting pathology-genomics imbalance yields a multimodal survival model that ranks patient risk more accurately than current multimodal baselines. Concretely, MRePath treats every patch of a whole slide image as a node in a hypergraph built from two kinds of hyperedges: topological ones connecting spatial neighbors within a distance threshold, and feature-based ones connecting patches with similar embeddings. A sheaf hypergraph Laplacian propagates information through these higher-order connections, producing pathology features that retain contextual and hierarchical detail. In parallel, a dynamic weighting scheme computes mono-confidence (per-modality reliability) and holo-confidence (interaction-aware reliability) to reweight pathology and genomic features, and interactive co-attention layers align the two modalities for the final hazard prediction. The reported result is a mean C-Index of 71.5% over BLCA, BRCA, CO-READ, HNSC and STAD, outperforming the second-best compared method by roughly 3 percentage points.","pith_inferences":["A natural extension, hinted at in the limitations, is to apply the mono/holo-confidence weighting to other modality pairs with very different dimensionalities, such as imaging paired with structured clinical records; the mechanism does not depend on the specific biology of pathology or genomics.","The contextual-detail claim could be isolated by ablating the hypergraph to a k-nearest-neighbour graph while keeping the same number of neighbours; the authors do not report this comparison, and it would test whether the gains come from higher-order structure or simply from broader local context.","The acknowledged equal-k limitation suggests an adaptive per-slide threshold, for example scaling k with patch count or tissue area, as a testable variant; matching or exceeding the reported 71.5% with a data-dependent threshold would weaken concerns that the fixed k=9 was tuned to the benchmark."],"forward_implications":["On all five cohorts, MRePath reports the highest C-Index, so the paper claims that hypergraph pathology features combined with rebalanced fusion improve risk ranking over unimodal and multimodal baselines alike.","The ablation study reports that combining topological and feature hyperedges (71.5%) beats either type alone (68.7% and 68.8%), implying the two sources of higher-order structure carry complementary prognostic signal.","The paper reports that dynamic weighting (71.5%) outperforms fixed weights (67.8%–69.5%) and that interactive alignment fusion beats single-direction cross-attention (66.9%–68.9%), supporting its diagnosis of pathology-genomics imbalance.","Kaplan-Meier splits by median predicted risk reach log-rank p<0.05 on all five datasets, which the paper reads as evidence that the model's risk scores stratify patients meaningfully."],"supporting_citations":[{"why":"Supplies the hypergraph neural network convolution that MRePath uses as its base aggregator for higher-order patch relationships.","marker":"[Feng et al., 2019]"},{"why":"Supplies the sheaf hypergraph Laplacian that replaces the ordinary Laplacian to enable structured information flow between nodes and hyperedges.","marker":"[Duta et al., 2024]"},{"why":"Defines the multimodal co-attention transformer (MCAT), the baseline whose evaluation protocol, dataset splits, and feature preprocessing the paper reuses.","marker":"[Chen et al., 2021b]"},{"why":"Provides MOTCat, one of the two strongest multimodal baselines that MRePath claims to outperform in the head-to-head comparison.","marker":"[Xu and Chen, 2023]"},{"why":"Provides PIBD, the strongest compared baseline, and the source for several unimodal and SNN+CLAM results reproduced in the comparison table.","marker":"[Zhang et al., 2024]"},{"why":"Provides SurvPath, a dense-interaction multimodal baseline, and part of the evaluation setup the paper follows.","marker":"[Jaume et al., 2024]"},{"why":"Supplies attention-based MIL (ABMIL), the patch-aggregation approach the paper argues loses contextual and hierarchical detail.","marker":"[Ilse et al., 2018]"},{"why":"Supplies the negative log-likelihood survival loss used to train the models in the paper.","marker":"[Yao et al., 2020]"},{"why":"Provides the model-confidence learning idea on which the mono-confidence measure is based.","marker":"[Corbière et al., 2019]"}],"fun_headline_variants":["Rebalanced hypergraph fusion raises survival C-Index to 71.5%","Hypergraph context beats patch aggregation for cancer prognosis","Cross-modality rebalance sharpens risk ranking in five cancers","Higher-order pathology features lift multimodal survival scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hyperedge construction threshold k=9, chosen after inspecting C-Index results on the full benchmark, generalizes to unseen cross-validation folds; if that choice was tuned on the test folds rather than nested within training folds, part of the reported gain may reflect overfitting rather than genuine improvement.","fun_headline_variants_meta":{"raw":{"variants":["Rebalanced hypergraph fusion raises survival C-Index to 71.5%","Hypergraph context beats patch aggregation for cancer prognosis","Cross-modality rebalance sharpens risk ranking in five cancers","Higher-order pathology features lift multimodal survival scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00089,"raw_usage":{"total_tokens":3820,"prompt_tokens":906,"completion_tokens":2914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2845}},"tokens_in":522,"tokens_out":2914,"duration_ms":19032,"temperature":1.0,"reasoning_tokens":2845,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:43:18.988095+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same 5-fold protocol with k selected per training fold, either nested or on a held-out validation split, and compare the mean C-Index against PIBD and MOTCat; if the 3-point gap shrinks to within one standard deviation, the central superiority claim is not supported.","supporting_citations":[{"cited_title":"Hypergraph neural net- works","cited_arxiv_id":null,"evidence_quote":"Supplies the hypergraph neural network convolution that MRePath uses as its base aggregator for higher-order patch relationships."},{"cited_title":"Sheaf hypergraph networks","cited_arxiv_id":null,"evidence_quote":"Supplies the sheaf hypergraph Laplacian that replaces the ordinary Laplacian to enable structured information flow between nodes and hyperedges."},{"cited_title":"Prototypical information bottlenecking and disentangling for multimodal cancer survival prediction","cited_arxiv_id":null,"evidence_quote":"Provides PIBD, the strongest compared baseline, and the source for several unimodal and SNN+CLAM results reproduced in the comparison table."},{"cited_title":"Modeling dense multimodal in- teractions between biological pathways and histology for survival prediction","cited_arxiv_id":null,"evidence_quote":"Provides SurvPath, a dense-interaction multimodal baseline, and part of the evaluation setup the paper follows."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"Supplies attention-based MIL (ABMIL), the patch-aggregation approach the paper argues loses contextual and hierarchical detail."},{"cited_title":"Whole slide images based cancer survival prediction using at- tention guided deep multiple instance learning networks","cited_arxiv_id":null,"evidence_quote":"Supplies the negative log-likelihood survival loss used to train the models in the paper."}],"review_version":1}