{"id":"ffe83175-0b89-4cbb-8e06-dae6667d03b9","arxiv_id":"2506.19324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"M2Surv reports higher concordance index than prior multimodal survival models on five TCGA datasets by adding hypergraph-based multi-slide fusion, gene-attention hyperedges, and a memory bank for missing-modality inference.","lead":"This paper presents M2Surv, a cancer survival model that combines several pathology slide images with gene data using hypergraph networks and a memory bank. It reports small C-index gains on five TCGA datasets and shows the memory bank helps when one data type is missing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The incomplete-modality claim is only tested for test-time missingness after complete-pair training; patients with missing modalities during training—the motivating scenario—have no paired memory entries to retrieve, so the 3.3%/7.9% gains are not yet established for genuine incomplete cohorts.","rationale":"I agree with the reader's weakest_assumption; it is the most load-bearing soft spot. The full-modality SOTA claim is supported by the reported mean C-indexes, though significance is underreported; the incomplete-modality claim is the paper's distinctive contribution and is evaluated in an easier regime than the one the introduction motivates. The proposed simulation directly tests whether memory retrieval works when training data themselves contain missing modalities. If it does not, the 3.3%/7.9% margins would not generalize to the motivating clinical setting. This does not overturn the full-modality results, so the CONDITIONAL verdict stands as UNCHANGED.","tokens_in":9464,"tokens_out":9882,"duration_ms":116144,"concrete_test":"Run a missing-at-training experiment on BLCA (or all five datasets): randomly mask, say, 25% and 50% of training patients' genomic profiles (and, in a second arm, their WSIs) before any memory-bank construction. For each masked-at-training patient, report C-index (i) using only the observed modality and (ii) using the memory bank to retrieve the missing modality from the remaining complete-pair patients. Compare with the full-pair-training results reported in Table 1 and Fig. 2c. If (ii) does not exceed the strongest unimodal baseline in that masked condition, the central incomplete-modality claim is limited to test-time-only missingness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the abstract's incomplete-modality claim to hold, the memory bank must compensate for missing modalities in a clinically realistic cohort. Section 2.4 populates the memory bank only from complete pathology–genomic pairs seen during training, and Section 3.2/Fig. 2c evaluates incompleteness by dropping a modality at inference after training on complete pairs. If missing modalities are present in the training cohort—the scenario the introduction motivates via insufficient tissue and sequencing errors—those patients cannot contribute paired features to the memory bank, and the retrieval step (Eq. 5) has no stored cross-modal associations to draw on. The paper never simulates missing-at-training, so the reported 3.3% pathology-only and 7.9% gene-only advantages may only reflect access to paired historical features at inference, not robust handling of actual incomplete cases. The Limitations section (Section 4) concedes that retrieval depends on training-data coverage, which is precisely the untested condition here.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes M2Surv, a multimodal survival prediction framework that combines a multi-slide hypergraph for pathology features, a gene-attentive hypergraph for cross-modal fusion, and a momentum-updated memory bank that retrieves previously seen paired pathology-genomic features to compensate for missing modalities at inference. The method is evaluated on five TCGA datasets with 5-fold cross-validation and C-Index as the metric, reporting a mean C-Index of 0.701 versus 0.685 for the strongest prior baseline and 3.3% and 7.9% gains in pathology-only and gene-only incomplete-modality scenarios, respectively.","tokens_in":9640,"tokens_out":5394,"duration_ms":60504,"significance":"If the reported results hold, M2Surv is a practically useful contribution to multimodal survival prediction, particularly for combining multiple FFPE and FF slides and for mitigating modality imbalance through gene-attentive hyperedges. The memory-bank design is lightweight and is also shown to be adaptable to other multimodal models, and the authors release code. These are genuine strengths. However, the empirical support for the headline claims is currently incomplete: the main comparison lacks significance testing, and the incomplete-modality evaluation does not cover the clinically motivating scenario of modalities missing during training, which is precisely the condition on which the memory mechanism depends.","major_comments":[{"comment":"The headline improvement is not statistically established. The absolute mean C-Index gain over the strongest baseline PIBD is 0.016 (from 0.685 to 0.701), i.e., the abstract's \"over 2.3%\" is a relative rather than absolute improvement, and this is not stated. On most datasets the reported differences from several baselines are within one standard deviation, and no significance tests, confidence intervals, or paired comparisons are provided. In addition, some baselines are quoted from [29] rather than reproduced under the same folds and preprocessing, so the comparison may be confounded. Please provide per-dataset significance testing with all baselines evaluated under an identical protocol, or substantially temper the superiority claim.","section":"Table 1, Sec. 3.1"},{"comment":"The incomplete-modality evaluation only tests missingness at inference after training on complete pathology-genomic pairs. The memory bank stores only paired features from complete training samples, so a patient whose modality is missing during training cannot contribute a cross-modal entry for the retrieval in Eq. (5). This is exactly the scenario motivated in the Introduction (insufficient tissue, sequencing errors), yet the paper never simulates missing-at-training modalities. The Limitations paragraph in Sec. 4 concedes that retrieval efficacy depends on training-data coverage, which is the untested condition. The 3.3% pathology-only and 7.9% gene-only claims are therefore unvalidated for the motivating setting; a missing-at-training experiment, or a clear restriction of the claims, is required.","section":"Sec. 2.4, Sec. 3.2, Fig. 2c"},{"comment":"Several key hyperparameters are selected using the same datasets whose test performance is then reported as the final result. The hyperedge construction threshold λ is selected from the ablation table (5, 9, 25), the gene-attentive threshold β is described as \"empirically set\" to the top 5%, and the retrieval count μ is chosen as the best value in Fig. 2b. Without nested validation or a separate model-selection protocol, this selection on the evaluation data can inflate the reported mean C-Index. Please describe exactly how tuning was performed and provide sensitivity results at the selected hyperparameter values.","section":"Sec. 3.2, Table 2, Fig. 2b"},{"comment":"The rows labeled \"M²Surv (Ours)\" in the Pathology and Genomic blocks are not comparable to the unimodal baselines in the same blocks. These rows correspond to a model trained on both modalities with one modality removed at inference and compensated by the memory bank, whereas ABMIL, TransMIL, MLP, and SNN are single-modality models trained from scratch. The comparison therefore conflates the effect of the memory mechanism with the quality of the unimodal representation. Please include a genuinely unimodal M2Surv variant, or label these rows explicitly as modality-dropout experiments rather than unimodal comparisons.","section":"Table 1, Pathology and Genomic rows"}],"minor_comments":[{"comment":"The phrase \"Fresh Froze\" should read \"Fresh Frozen\".","section":"Abstract and Introduction"},{"comment":"The distance condition is written as \"ζpj − ζpk ∥2 ≤ δ\"; a norm delimiter is missing in the typeset expression.","section":"Eq. (2)"},{"comment":"The header \"F use\" appears to mean \"Fusion\", and row labels such as \"HGNNM ulti9\" contain spacing artifacts that should be corrected.","section":"Table 2"},{"comment":"The definition of c_i is unconventional (c_i = 0 indicates event occurrence); although the loss terms are consistent with this coding, the meaning of \"event status\" should be stated explicitly to avoid confusion.","section":"Eq. (1)"},{"comment":"The text uses \"MCat\" in the memory-bank experiment while Table 1 and the main text use \"MCAT\"; please standardize the notation.","section":"Fig. 2c and Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version first: this is a workmanlike extension of the authors' own hypergraph survival line, with a memory bank bolted on for missing-modality inference. The empirical gains are real but modest, and the incomplete-modality claim is only half-tested. The paper is worth a serious look, but not a slam dunk.\n\nWhat's new: the multi-slide hypergraph construction (intra-slide spatial hyperedges plus inter-slide structural hyperedges) is a sensible way to handle FF/FFPE slide heterogeneity, and the gene-attentive hypergraph does a reasonable job of mitigating the pathology-genomics imbalance by giving every gene group dense connections to patches. The memory bank is a simple, cheap mechanism—4.5% extra training time, 2.5% extra inference—and it does help when a modality is dropped at test time. The ablation study is actually informative: they compare aggregators, hyperedge types, thresholds, and fusion methods, and the choices are not arbitrary.\n\nThe soft spots are in proportion. The stress-test note is on target: the incomplete-modality scenario only covers inference-time missingness after training on complete pairs. If a tumor sample lacks genomic data at training—the exact scenario the introduction motivates via sequencing errors—there are no paired features in the bank to retrieve, and the 3.3% and 7.9% gains disappear. The paper never simulates missing-at-training. The own Limitations section concedes that retrieval depends on training-data coverage, which is precisely the untested condition.\n\nAlso, most per-dataset differences are within one standard deviation, and no significance tests are reported. Baselines are a mix of quoted numbers from PIBD and reproduced numbers; that's a protocol inconsistency. Hyperparameters like λ, β, and µ were selected after seeing the test data, so the reported numbers are a bit cherry-picked. These are fixable, but they keep the paper at \"conditional accept\" level, not \"state of the art.\"\n\nThe memory bank is not circular—it retrieves features, not labels—so that criticism doesn't land. The self-citation of [20] is fine; the contribution is incremental but clearly stated.\n\nBottom line: if you work on computational pathology or missing-modality fusion, this is worth a half-hour read. As a referee, I'd ask for a missing-at-training simulation and proper significance testing. Send it to review, but with those requests.","headline":"Solid incremental engineering; the incomplete-modality claim is only tested for test-time missingness, so the headline is somewhat overstated.","tokens_in":10209,"tokens_out":2495,"would_cite":false,"duration_ms":26475,"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":"A hypergraph-plus-memory framework lifts multimodal cancer survival prediction past prior methods.","keywords":["survival prediction","whole slide images","genomics","hypergraph learning","incomplete modality","memory bank","modality imbalance","cancer prognosis"],"falsifier":"Train the same pipeline on cohorts where a fixed fraction of patients lacks one modality during training, then evaluate on a complete test set: if the memory-augmented model no longer beats the best unimodal baseline, the paper's central claim about incomplete modalities is falsified. A simpler observable check is that the reported 3.3% and 7.9% gaps should reproduce when the bank is seeded with complete training pairs.","tokens_in":9229,"feed_emoji":"🧬","tokens_out":9875,"duration_ms":94639,"temperature":0.7,"pith_summary":"This paper tries to establish that cancer survival prediction improves when pathology slides and genomic data are fused through hypergraph structures rather than standard cross-attention, and that a memory bank can stand in for a missing modality at prediction time. The proposed model, M²Surv, treats patches from multiple slides as nodes in a hypergraph, capturing spatial structure inside each slide and shared morphology across slides, then links six gene groups densely to the pathology patches they attend to most. On five large public cancer cohorts, the authors report a mean concordance index (C-index, the standard ranking metric for survival models) of 0.701, ahead of prior multimodal methods by more than 2.3%. When either slides or genes are withheld at inference, the memory bank retrieves the closest complete patient pair and reports gains of 3.3% over pathology-only models and 7.9% over gene-only models. The result matters because real clinical records often have one missing assay, and most existing methods require both modalities.","feed_headline":"Hypergraphs plus memory bank lift multimodal survival prediction","feed_subtitle":"Slides and gene data are fused through hyperedges; missing data is filled by retrieving the closest stored patient.","key_machinery":"The load-bearing object is a pair of hypergraphs plus a memory bank. A hypergraph is a graph whose edges can link more than two nodes; here the nodes are pathology patches, plus six gene groups in the second stage. The multi-slide hypergraph builds intra-slide hyperedges from spatial neighborhoods and inter-slide hyperedges from patch-feature similarity, so both slide-specific staining variation and shared morphology are encoded. The gene-attentive hypergraph creates one dense hyperedge per gene group to its top-attending patches, which the paper argues rebalances the pathology-dominated fusion. The memory bank uses a momentum update during training and cosine-similarity retrieval with top-µ aggregation at inference to synthesize the missing modality. These components jointly carry the claim that performance gains come from higher-order slide interactions and balanced cross-modal fusion, not from a single architectural trick.","core_discovery":"The paper's central claim is that M²Surv improves pathology-genomics survival prediction by replacing single-slide, cross-attention fusion with two hypergraph stages: a multi-slide hypergraph that first links spatially neighboring patches inside each slide and then links similar patches across slides, and a gene-attentive hypergraph that connects each of six gene groups to the pathology patches it attends to most. This second stage is the paper's answer to modality imbalance: instead of letting thousands of patches swamp a few hundred genes, each gene group reaches every relevant patch through dense hyperedges. A momentum-updated memory bank stores paired pathology-genomic features from complete training patients; at inference, if one modality is missing, the available modality retrieves the top matching stored pair and uses it to approximate the absent one. On five large cancer cohorts the authors report a mean C-index of 0.701, surpassing prior multimodal methods by more than 2.3%, and in missing-modality settings beating pathology-only models by 3.3% and gene-only models by 7.9%.","pith_inferences":["The paper does not simulate missing modalities during training; a natural extension is to test whether the memory bank can be updated with partial pairs, since hospitals often lack one assay for some patients.","Because the best retrieval uses only the single most similar stored patient (µ=1), the quality of the approximation is bounded by how well the training cohort covers the test patient's biological profile, so rare subtypes are where this design is most likely to fail.","The dense gene-to-patch hyperedges could be reinterpreted as a weak form of cross-modal attention with an explicit balance mechanism; comparing against a loss function that directly penalizes modality imbalance would isolate where the improvement truly comes from.","If the multi-slide gain is driven mostly by fresh-frozen slides, archives that contain only formalin-fixed slides may see smaller benefits; the paper does not break down its 0.701 mean by slide type, so this boundary is untested."],"forward_implications":["Multi-slide hypergraph aggregation should replace single-slide pooling in multimodal survival models, because the ablations tie each hyperedge type (intra-slide and inter-slide) to a measurable C-index improvement.","The memory bank is a cheap, generalizable remedy for missing modalities: it adds about 4.5% training time and 2.5% inference time and restores near-full performance for the two co-attention baselines it was added to.","A mean C-index of 0.701 across five cohorts, exceeding prior multimodal methods by more than 2.3%, implies the design transfers across tumor types rather than overfitting one site.","Gene-attentive hyperedges make the model's cross-modal reasoning inspectable, since the attention scores highlight which gene groups target which tissue regions."],"supporting_citations":[{"why":"Supplies the class-wise memory bank concept that the paper adapts to store paired pathology-genomic features.","marker":"[1]"},{"why":"The co-attention transformer baseline whose genomic feature grouping and multimodal fusion M²Surv extends and must beat.","marker":"[3]"},{"why":"The hypergraph convolution used to update node representations in both hypergraph stages.","marker":"[8]"},{"why":"The momentum contrast update rule adopted for refreshing stored memory-bank features.","marker":"[10]"},{"why":"The dense pathway-histology multimodal baseline that motivates the gene-attentive hypergraph design.","marker":"[12]"},{"why":"The attention-based patch aggregator that provides the 4096-patch extraction protocol and a pathology-only baseline.","marker":"[16]"},{"why":"A transformer-based multiple-instance pathology baseline compared in the incomplete-modality experiments.","marker":"[22]"},{"why":"An optimal-transport co-attention baseline compared in the main results and used to test memory-bank adaptability.","marker":"[26]"},{"why":"The prototypical information bottleneck baseline that supplies quoted unimodal results, dataset splits, and comparison settings.","marker":"[29]"}],"fun_headline_variants":["Memory bank fills missing slides or genes for survival prediction","Hypergraph and memory tackle incomplete cancer data","M2Surv: memory-augmented hypergraph improves survival prediction","Survival prediction with incomplete modalities via hypergraph memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The memory bank is built only from patients who have both slide and gene data during training, so the method's value for truly incomplete datasets, where some patients are missing data from the start, remains unshown.","fun_headline_variants_meta":{"raw":{"variants":["Memory bank fills missing slides or genes for survival prediction","Hypergraph and memory tackle incomplete cancer data","M2Surv: memory-augmented hypergraph improves survival prediction","Survival prediction with incomplete modalities via hypergraph memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2869,"prompt_tokens":985,"completion_tokens":1884,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":1819}},"tokens_in":601,"tokens_out":1884,"duration_ms":15695,"temperature":1.0,"reasoning_tokens":1819,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:07:40.493882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline on cohorts where a fixed fraction of patients lacks one modality during training, then evaluate on a complete test set: if the memory-augmented model no longer beats the best unimodal baseline, the paper's central claim about incomplete modalities is falsified. A simpler observable check is that the reported 3.3% and 7.9% gaps should reproduce when the bank is seeded with complete training pairs.","supporting_citations":[{"cited_title":"In: ICCV","cited_arxiv_id":null,"evidence_quote":"Supplies the class-wise memory bank concept that the paper adapts to store paired pathology-genomic features."},{"cited_title":"In: AAAI","cited_arxiv_id":null,"evidence_quote":"The hypergraph convolution used to update node representations in both hypergraph stages."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"The momentum contrast update rule adopted for refreshing stored memory-bank features."},{"cited_title":"In: CVPR","cited_arxiv_id":null,"evidence_quote":"The dense pathway-histology multimodal baseline that motivates the gene-attentive hypergraph design."},{"cited_title":"Nature biomedical engineer- ing5(6), 555–570 (2021)","cited_arxiv_id":null,"evidence_quote":"The attention-based patch aggregator that provides the 4096-patch extraction protocol and a pathology-only baseline."},{"cited_title":"NeurIPS34, 2136–2147 (2021)","cited_arxiv_id":null,"evidence_quote":"A transformer-based multiple-instance pathology baseline compared in the incomplete-modality experiments."},{"cited_title":"In: ICLR (2024)","cited_arxiv_id":null,"evidence_quote":"The prototypical information bottleneck baseline that supplies quoted unimodal results, dataset splits, and comparison settings."}],"review_version":1}