{"id":"f1d29cee-18e4-4306-a060-add5fe978842","arxiv_id":"2411.11487","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Grouping multiple whole-slide images into a single sample and scanning them with a Mamba-based graph-sequence model improves survival prediction C-index on five TCGA datasets.","lead":"This paper introduces GroupMIL, a survival prediction framework that treats several pathology slides from the same cancer type as one group instead of analyzing each slide alone. The authors report higher concordance indices than existing methods on five TCGA cancer datasets and argue that cross-slide context improves prognosis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Group construction is under-specified with respect to cross-validation and patient boundaries; if validation groups can include training-fold slides, the reported 0.737 mean C-index may be inflated by inter-slide leakage.","rationale":"The single most load-bearing concern is the integrity of the empirical comparison, because the central claim is entirely based on the Table 1 C-index advantage. The novelty of modeling groups of slides makes group construction the critical protocol detail: if groups are not confined within cross-validation folds and patient/case boundaries, then the inter-slide interactions in GPAMamba can leak training information into validation predictions, directly inflating the reported mean C-index. This is not an internal inconsistency in the architecture; it is an evaluation-protocol ambiguity that must be resolved before any of the downstream results, ablations, or clinical-consistency analyses can be interpreted. The reader's weakest assumption identifies exactly this issue, and I agree with it; I would sharpen it by noting that TCGA cases often contain multiple WSIs, so a slide-level rather than patient-level split would itself cause leakage even if folds were respected. Other weaknesses, such as missing significance tests and ablation deltas smaller than fold noise, are real but secondary: they matter only once the comparison protocol is clean. The proposed concrete test directly checks whether the protocol permits training/validation mixing and whether the reported margin survives a leakage-free rerun. If it does not survive, the central claim should be substantially weakened; if it does survive, the conditional concerns about this specific issue are resolved, though significance testing would still be advisable. Since the reader already issued a CONDITIONAL verdict based on this concern, my stress-test does not move the verdict; it reinforces the same condition with a sharper patient-level formulation and a precise quantitative threshold for what would falsify the headline claim.","tokens_in":13252,"tokens_out":6509,"duration_ms":76378,"concrete_test":"Obtain or reconstruct the data loader and fold split used for Table 1, then verify two invariants: (1) the cross-validation split is at the patient/case level, not the slide level, and (2) every validation group consists exclusively of slides from the validation fold, with the same holding for training groups. Then rerun the five-fold experiment with groups randomly re-sampled inside each fold across five seeds. If any submitted validation group contains a slide from a training fold, or if the leakage-free mean C-index drops by more than 0.027 relative to Table 1, the claimed advantage over MambaMIL is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Sec. 4.4, Table 1) is that GroupMIL 'significantly outperforms' state-of-the-art survival models, but the evaluation protocol does not state how slide groups are formed relative to patient and fold boundaries. Sec. 3.2 defines a group as 'every B random WSIs' of the same cancer type, and Sec. 4.3 says only that 'during model training and validation, we treat every 6 slides as a group.' Nothing in the text says that groups are sampled separately inside each cross-validation fold, or that the underlying split is at the patient/case level rather than the slide level. If groups are formed by random assignment over all same-cancer slides, then a validation group can contain slides from training folds. Because GPAMamba's inter-slide interactions (Alg. 1, Sec. 3.3) propagate information across slides in the group before the dual-head predictor emits a per-slide risk, a held-out slide's prediction can be influenced by training-fold slides whose survival labels were used to fit the model. This would inflate the C-index relative to any deployable single-patient protocol. The mean gain over MambaMIL is 0.027 (Table 1), and even the favorable comparison in Table 4 uses group sizes that could mix folds. With no code and no significance tests, the 'significantly outperforms' claim is not yet supported unless group construction is constrained and disclosed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes GroupMIL, a survival prediction framework that departs from single-slide processing by modeling a group of whole-slide images as a single sample. The method introduces PAMamba for long patch-sequence modeling, GPAMamba for alternating intra-slide graph convolution and inter-slide sequence scanning, and a dual-head predictor combining Cox and discrete-time losses. Experiments on five TCGA cancer datasets report mean C-index values around 0.737, outperforming ten baseline methods, with additional ablation studies, attention visualizations, and Kaplan-Meier stratification analyses. The central claim is that GroupMIL significantly outperforms state-of-the-art approaches.","tokens_in":13582,"tokens_out":2857,"duration_ms":29372,"significance":"If the reported results hold under a rigorously controlled protocol, the group-based modeling idea is a meaningful conceptual contribution to computational pathology: it is clinically motivated, technically novel in combining Mamba with graph convolutions, and the dual-head predictor is a sensible integration of two standard survival losses. The paper also includes useful ablations and a discussion that candidly raises the question of whether gains are merely due to larger batch size. The main value is in opening a new axis of modeling—cross-slide context—that could generalize to other whole-slide-image tasks. However, the significance is currently contingent on the evaluation protocol being leak-free and statistically sound, which the manuscript does not yet establish.","major_comments":[{"comment":"The group construction protocol is under-specified with respect to cross-validation and patient boundaries. Section 3.2 defines a group as 'every B random WSIs' of the same cancer type, and Section 4.3 states only that 'during model training and validation, we treat every 6 slides as a group.' The manuscript never states whether groups are formed separately within each fold, whether the underlying split is at the patient/case level rather than the slide level, or whether a patient's slides can appear in both training and validation groups. Because GPAMamba propagates information across slides within a group (Algorithm 1, lines 5-11) before the predictor emits per-slide risk, a validation slide whose group contains training-fold slides could have its prediction influenced by slides whose survival labels were used to fit the model. This would inflate the reported C-index gains in Table 1 relative to any deployable single-patient protocol. The authors must specify the grouping procedure relative to folds and patients, and re-run the evaluation with groups constrained to be within each validation fold and with patient-level splits if not already used.","section":"Sec. 4.3 and Sec. 3.2"},{"comment":"The claim that GroupMIL 'significantly outperforms' state-of-the-art approaches is not supported by any statistical significance test. Table 1 reports only the mean and standard deviation of the C-index over five folds, and for several datasets the standard deviations overlap between GroupMIL and the best baseline (e.g., BRCA: 0.713 ± 0.052 vs. MambaMIL 0.654 ± 0.042; UCEC: 0.760 ± 0.047 vs. 0.743 ± 0.055). With five folds, such overlapping intervals cannot justify the word 'significantly.' The authors should add paired significance tests (e.g., Wilcoxon signed-rank or permutation tests across folds) or temper the claim in the abstract and Section 4.4.","section":"Abstract and Sec. 4.4, Table 1"},{"comment":"All architectural and hyperparameter choices—group size, number of graph neighbors, loss weight α, gradient accumulation steps, and interval count—are selected based on the same five TCGA datasets used for the final evaluation. This creates a risk of selection bias and optimistic performance estimates, since the model is tuned on the test data. The ablation tables (Tables 2 and 3) report performance for different choices but do not use nested cross-validation or a held-out tuning set. The authors should either provide a nested cross-validation analysis or explicitly state that the reported numbers are tuning-selected and therefore should be interpreted as potentially optimistic.","section":"Sec. 4.5 and Sec. 5"}],"minor_comments":[{"comment":"The reported standard deviation for TransMIL on UCEC is 0.751, which is implausibly large compared to all other entries in the table and is likely a typo for 0.071 or 0.075; please correct.","section":"Sec. 4.4, Table 1"},{"comment":"The sentence 'To harness the rich information withing′' contains a typo: it should be 'within g′'.","section":"Sec. 3.4"},{"comment":"The symbol B is used to denote group size without formal definition in Section 3.2; it is only later clarified as 6 in Section 4.3. Please define B when first introduced.","section":"Sec. 3.2"},{"comment":"The notation in Eq. (3), ϕ(x) = S(σ(C(L(x)))) + σ(L(x)), is confusing because L, C, σ, and S are not dimensionally specified. Please clarify the shapes and the order of operations, or add a sentence describing the role of each operator.","section":"Sec. 3.3, Eq. (3)"},{"comment":"The batch-size/group-size comparison in Table 4 is presented without any significance testing; the conclusion that 'there is no strong correlation between batch size and model performance' is based on overlapping means and standard deviations. Please add a statistical comparison or soften the claim.","section":"Sec. 5, Table 4"},{"comment":"Some references appear incomplete or informal, e.g., reference [13] cites 'Mamba: Linear-time sequence modeling with selective state spaces' as an arXiv preprint without a year in the citation. Please verify all references for completeness.","section":"Sec. 2, References"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a computational pathology or medical imaging journal and the core idea is interesting. However, the evaluation protocol is currently not rigorous enough to support the central claim: the group formation process is ambiguous with respect to patient and fold boundaries, and no statistical significance testing accompanies the 'significantly outperforms' assertion. I would encourage the editors to require the authors to release code and a detailed data-splitting protocol, and to conduct a leakage-controlled re-evaluation. If the gains persist under a clean protocol, this could be a solid contribution; otherwise the present evidence is insufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this as a promising new idea in computational pathology with an unproven headline claim. The novelty is real: instead of treating each WSI as an independent sample, they group several slides into one sample and run GPAMamba, which alternates graph convolution with dual-branch Mamba scanning over the long patch sequence and over slide-level representations. The empirical gain over MambaMIL is a mean C-index of 0.737 vs 0.710 across five TCGA datasets, and the group-level direction is worth taking seriously.\n\nWhat the paper does well: the batch-size control experiment in Table 4 directly addresses the obvious confound that grouping is just a bigger effective batch. Even after that control, grouping helps most models and their own model most clearly. The discussion section is candid, listing open questions about how to form groups, and the ablations are fairly thorough. The KM curves and stage-correlation analyses are a useful bonus, even if they are secondary.\n\nThe soft spots are real. Most importantly, the grouping protocol is under-specified. Section 3.2 says \"every B random WSIs\" and Section 4.3 says \"we treat every 6 slides as a group,\" but nothing states whether groups are sampled separately inside each fold or whether a patient's slides can appear in both training and validation groups. Given that GPAMamba propagates information across slides in a group before prediction, a validation slide could be influenced by training-fold slides if groups mix folds. That would inflate the C-index. The \"significantly outperforms\" language is also unsupported: no significance tests, no confidence intervals, and several fold standard deviations overlap. The lack of code compounds this. Minor points: the UCEC value for TransMIL in Table 1 (0.711 ± 0.751) looks like a typo, and all hyperparameters were selected on the same five datasets, so the comparison is slightly optimistic.\n\nWho is this for? Researchers in computational pathology and MIL, particularly those working on cross-slide context and on evaluation hygiene in slide-level models. It deserves a serious referee: the core idea is novel, the paper is honest about its limitations, and the issues are fixable. I would accept it for peer review but with the expectation of major revision: disclose the grouping protocol precisely, add significance testing, and soften the central claim or ship code and evidence that the reported gain survives proper patient-level grouping.","headline":"Promising group-level modeling idea with an honest discussion, but the 'significantly outperforms' claim is not supported by the statistics, and the under-specified grouping protocol could leak across cross-validation folds.","tokens_in":14075,"tokens_out":2434,"would_cite":true,"duration_ms":26923,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Modeling a group of six whole-slide images as one sample improves cancer survival prediction, with a reported mean concordance index of 0.737 across five TCGA datasets.","keywords":["survival prediction","whole slide images","multiple instance learning","state space models","graph neural networks","cross-slide modeling","computational pathology","TCGA"],"falsifier":"A re-run of the five-fold evaluations in which every group is built from slides belonging to a single fold, and no patient's slides appear in both a training and a validation group, would settle the claim: if the mean C-index falls back to the one-slide baseline, the reported 0.737 is inflated by cross-split leakage rather than produced by cross-slide modeling.","tokens_in":13088,"feed_emoji":"🔬","tokens_out":9579,"duration_ms":86123,"temperature":0.7,"pith_summary":"This paper tries to establish that survival prediction from pathology slides improves when a model looks at a group of slides from several patients at once instead of each slide alone. It introduces GroupMIL, which treats a random group of six same-cancer whole-slide images as one sample, and GPAMamba, which alternates graph-based local aggregation with long-sequence scanning to combine within-slide and cross-slide patterns. A dual-head predictor merges a risk score with discretized survival probabilities into one composite risk measure. On five TCGA datasets the paper reports a mean C-index of 0.737 versus 0.710 for MambaMIL, and it argues the improvement comes from cross-slide context rather than from simply using a larger batch.","feed_headline":"Grouped slides beat single-slide models in cancer survival prediction","feed_subtitle":"Reading six tumor slides at once lets each case borrow evidence from others, improving survival-prediction score to 0.737.","key_machinery":"The central object is GPAMamba, a block that combines graph convolutions with a position-and-attention scanning layer (PAMamba) built on Mamba, a linear-time state-space sequence model. In each block, every slide is a fixed graph whose nodes are patches and edges connect nearby patches; graph-convolution layers exchange local information. PAMamba then flattens the group's patches into one long sequence and scans it twice, once in spatial order and once reordered by learned prognostic-attention weights, before merging the two aligned scans. The sequence is split back into per-slide graphs and convolved again, an intra-inter-intra alternation. After graph pooling, a second Mamba module scans the ordered slide representations, and a dual-head predictor maps them to a risk score and an interval survival-probability vector. This machinery is what lets one slide's representation borrow evidence from other slides in the group.","core_discovery":"The paper's central claim is that survival prediction improves when a model reads several whole-slide images together. It argues that existing one-slide-at-a-time pipelines cannot capture slide-agnostic phenotypes, whereas a group exposes comparable prognostic features across patients. GroupMIL forms a group of six same-cancer slides, treats it as one sample, and runs GPAMamba, which alternates graph convolutions over each slide's patch graph with two long-sequence scans of the concatenated patches; the resulting slide representations are scanned again to add cross-slide context. The dual-head predictor then combines a risk score with discretized survival probabilities into a single composite risk. The paper reports a mean concordance index (C-index), a standard rank-agreement metric for survival predictions, of 0.737 over five TCGA datasets, compared with 0.710 for MambaMIL, and interprets the gain as evidence that collective analysis works.","pith_inferences":["Group modeling behaves like a transferable wrapper: the paper's batch-versus-group comparison shows several tested baselines improve when their slides are grouped, so the collective-analysis mechanism is not tied only to GPAMamba's specific internals.","Group composition is currently random and within one cancer type; grouping by age, stage, or molecular subtype is a natural next experiment, and the authors explicitly flag it as future work.","A deployment caution follows from the paper's silence on whether groups respect fold boundaries: any practitioner adopting group modeling should construct groups within patient-level splits to keep validation honest."],"forward_implications":["Survival models can stop treating slides as isolated samples; each patient's risk can be informed by comparable slides from other patients with the same cancer.","Combining a risk head with a discrete-time survival-probability head yields a composite risk score that the paper reports stratifies patients significantly on all five datasets.","Attention orderings in the model highlight tumor and necrotic patches over normal stroma and muscle, giving pathologists a visual check consistent with their own assessment.","Because the Mamba-based long-sequence module is reported to beat a comparable Transformer with lower FLOPs and memory, group-scale patch sequences remain computationally feasible.","The group-size ablation shows stable performance for groups of 4, 6, or 8 slides, so the framework is not tightly tuned to exactly six slides."],"supporting_citations":[{"why":"Supplies the linear-time state-space operator that PAMamba and the slide-group module are built on.","marker":"[13]"},{"why":"Provides the strongest single-slide baseline at 0.710 mean C-index and the sequence-reordering strategy GroupMIL extends.","marker":"[38]"},{"why":"Defines the patch-graph survival baseline used in comparisons and the conversion of survival probabilities into risk used in the dual-head score.","marker":"[2]"},{"why":"Serves as the transformer-based multiple-instance-learning baseline in the main comparison and group-generalization experiments.","marker":"[31]"},{"why":"Serves as the graph-CNN survival baseline and is among the models that improve when grouped in the batch-versus-group experiments.","marker":"[23]"},{"why":"Defines the partial-likelihood survival loss used by the risk head.","marker":"[5]"},{"why":"Defines the discrete-time survival loss used by the probability head and the interval-based probability output.","marker":"[41]"},{"why":"Supplies the patch segmentation and feature-extraction pipeline that produces the 1024-dimensional patch embeddings.","marker":"[26]"}],"fun_headline_variants":["Grouped slides boost cancer survival prediction to 0.737","Multi-slide analysis improves cancer survival prediction","GroupMIL: reading six slides at once improves prognosis","Cancer survival prediction gets boost from grouped slides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that forming random six-slide groups from the same cancer type, across training and validation, is a clean experiment rather than a leaky one; the paper does not state that groups are confined within folds or that a patient's slides never appear in both training and validation groups.","fun_headline_variants_meta":{"raw":{"variants":["Grouped slides boost cancer survival prediction to 0.737","Multi-slide analysis improves cancer survival prediction","GroupMIL: reading six slides at once improves prognosis","Cancer survival prediction gets boost from grouped slides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000738,"raw_usage":{"total_tokens":3275,"prompt_tokens":902,"completion_tokens":2373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2311}},"tokens_in":518,"tokens_out":2373,"duration_ms":16770,"temperature":1.0,"reasoning_tokens":2311,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:26:43.889683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A re-run of the five-fold evaluations in which every group is built from slides belonging to a single fold, and no patient's slides appear in both a training and a validation group, would settle the claim: if the mean C-index falls back to the one-slide baseline, the reported 0.737 is inflated by cross-split leakage rather than produced by cross-slide modeling.","supporting_citations":[{"cited_title":"Whole slide images are 2d point clouds: Context-aware survival prediction using patch-based graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Defines the patch-graph survival baseline used in comparisons and the conversion of survival probabilities into risk used in the dual-head score."},{"cited_title":"Transmil: Transformer based correlated multiple instance learning for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Serves as the transformer-based multiple-instance-learning baseline in the main comparison and group-generalization experiments."},{"cited_title":"Graph cnn for survival analysis on whole slide pathological images","cited_arxiv_id":null,"evidence_quote":"Serves as the graph-CNN survival baseline and is among the models that improve when grouped in the batch-versus-group experiments."},{"cited_title":"Regression models and life-tables","cited_arxiv_id":null,"evidence_quote":"Defines the partial-likelihood survival loss used by the risk head."},{"cited_title":"Bias in cross- entropy-based training of deep survival networks","cited_arxiv_id":null,"evidence_quote":"Defines the discrete-time survival loss used by the probability head and the interval-based probability output."}],"review_version":1}