{"id":"7e0605b6-d1c4-4aea-808d-916fedc04181","arxiv_id":"2607.27274","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A two-stage 'train features without labels, then aggregate at the patient level' framework—BridgeMIL—improves EEG disease classification accuracy in 14 of 15 dataset/backbone settings.","lead":"EEG diagnosis usually trains on short recording windows as if each window carried the patient's diagnosis. This paper instead learns window features with no labels first, then trains a patient-level model using only the patient's diagnosis, and reports higher accuracy on three EEG disease datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage 1 alignment may learn subject identity rather than disease-relevant structure; no direct test separates these, threatening the mechanism and generalizability of the accuracy gains.","rationale":"The reader's verdict is already CONDITIONAL, and the reader identified Stage 1's disease-relevance assumption as the weakest point. I agree that this is the single most load-bearing concern. The headline empirical claim is the accuracy advantage, but the scientific value of that claim depends on the mechanism: decoupled representation learning should capture diagnostic structure, not merely subject-specific nuisance. The Stage 1 objective explicitly encourages within-subject invariance, and VICReg prevents collapse without enforcing disease relevance. The paper's own visualization shows strong subject-wise clustering but only quantifies class separation. A linear-probe test directly separates these two hypotheses and would settle whether the mechanism holds. Other concerns (hyperparameter selection on ADFTD, no significance tests, confounded scarcity analysis) affect the strength of the evidence but do not cut as deeply at the central claim. Since the reader already recommended conditional acceptance and the concern can be addressed by an additional analysis, the verdict should remain UNCHANGED, with the concrete probe test as a recommended precondition for full acceptance.","tokens_in":18524,"tokens_out":4483,"duration_ms":54833,"concrete_test":"After Stage 1 pretraining on each dataset, freeze f_theta and train two linear probes—one predicting subject identity and one predicting disease label—using the same subject-level CV splits and training folds only. Compare balanced accuracies (or silhouette/ARI against subject vs. class labels) on held-out subjects. If the subject-ID probe achieves substantially higher accuracy than the disease probe (e.g., >90% vs. near chance), Stage 1 is dominated by subject-specific, non-diagnostic structure, and the decoupling claim is unsupported. If disease-probe accuracy is comparable to or exceeds subject-ID accuracy, the representations carry diagnostic information beyond subject identity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"BridgeMIL's central claim is that decoupling unsupervised instance representation learning from subject-level MIL supervision yields better subject-level EEG diagnosis. This rests on the Stage 1 premise that nearby-window and within-subject sub-bag alignments (Eqs. (2)-(4) and (9), Section 3.1) capture latent structure informative for the disease label. The objective pushes representations from the same subject to be similar (sub-bag alignment) and temporally nearby windows to be similar (nearby-window alignment). VICReg prevents collapse but does not select disease-relevant variance over subject identity, recording-condition drift, or non-pathological artifacts. The paper's own t-SNE (Figure 5, Section 4.4) shows distinct subject-wise clusters, and the representation metrics (SC, DB) are computed only with diagnostic classes, never with subject identity as the clustering target. If the dominant shared structure is subject identity, Stage 2's MIL aggregator can exploit subject-specific nuisance that correlates with diagnosis in these three small datasets (88, 63, 31 subjects), producing high in-dataset accuracy that may not generalize to new subjects or sites. This is load-bearing because it undermines both the claimed decoupling mechanism and the external validity of the 14/15 accuracy result. The ablation (Table 2) only shows that removing components hurts accuracy; it does not establish that the learned alignments are disease-relevant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses subject-level EEG disease diagnosis, arguing that inherited-label instance training and end-to-end MIL both have weaknesses. It proposes BridgeMIL: Stage 1 uses a VICReg-style objective to align nearby windows and within-subject sub-bags without any instance labels, producing a pretrained encoder; Stage 2 attaches an attention-based MIL aggregator, fine-tunes with subject-level cross-entropy, and adds an instance-level feature-retention loss that anchors the encoder to its Stage 1 representation. The method is evaluated on ADFTD, Mumtaz2017, and Rockhill2021 with five backbones, reporting the highest mean accuracy in 14 of 15 dataset-backbone settings and an overall mean accuracy of 76.57%. Additional analyses study inherited-label reliability, subject vs. instance scarcity, ablations, and representation structure.","tokens_in":18795,"tokens_out":6160,"duration_ms":66063,"significance":"If the empirical results hold, BridgeMIL is a practically useful two-stage recipe for small-cohort EEG diagnosis: it obtains a strong encoder from abundant unlabeled instances and preserves it during a lightly supervised MIL fine-tune. The experimental design has real strengths: subject-level stratified cross-validation is used consistently, the coverage of multiple backbones and datasets is commendable, the baselines are reasonably comprehensive, and the headline accuracy arithmetic is reproducible from Table 1. However, the central mechanistic claim—that the Stage 1 alignments capture disease-relevant structure rather than subject identity or recording-condition drift—is not directly established. Given the small subject counts (88, 63, 31), this limits the strength of the external-generalization conclusion. The paper would be substantially strengthened by a subject-identity control or a held-out-site experiment, a nested or fixed hyperparameter-evaluation statement, and a deconfounded scarcity analysis.","major_comments":[{"comment":"The Stage 1 objective (Eqs. (2)-(4), (9)) aligns nearby windows and same-subject sub-bags, so a representation that encodes subject identity or recording-condition drift would satisfy the objective as easily as one that encodes disease-relevant structure. VICReg prevents collapse but does not select disease-relevant variance over subject-identity variance. The paper's t-SNE (Fig. 5) actually displays distinct subject-wise clusters, and the SC/DB indices are computed only with diagnostic classes as targets; no subject-identity probe or clustering metric is reported. This is load-bearing for the central claim and for external validity: with 88, 63, and 31 subjects, subject-specific nuisance can correlate with diagnosis and inflate in-dataset accuracy. Please add a concrete control, e.g., a linear probe for subject identity versus diagnosis, a leave-one-site-out experiment, or a baseline th","section":"Section 3.1, Eq. (9), Figure 5"},{"comment":"The hyperparameters λ_ret and K are tuned on ADFTD with EEGNet and then used for all datasets and backbones in Table 1, without nested cross-validation or a clear statement that they were fixed before the main evaluation. Table 3 shows the adopted values are the best among those tested on ADFTD, which means the ADFTD column (and hence the overall 14/15 and 76.57% claims) may be optimistically biased. Please either report nested CV for hyperparameter selection, show the sensitivity results on the other two datasets, or explicitly document that the same hyperparameters were chosen a priori and held fixed; otherwise the main comparison is not a fully honest evaluation.","section":"Appendix F, Table 3"},{"comment":"The subject-scarcity experiment is confounded. Each condition draws a random subset of 8, 6, 4, or 3 folds from the 10 folds and then selects validation and test folds from that subset, so the test subject set changes across conditions. The accuracy differences in Fig. 4 therefore reflect both the number of training folds and the identity/composition of the test folds, violating the intended controlled comparison. To support the claim that subject scarcity is more limiting than instance scarcity, fix the test and validation folds across all conditions and vary only the number of training subjects; alternatively use random-subject subsampling within a fixed split. The instance-scarcity arm is not affected by this issue.","section":"Section RQ3, Figure 4"},{"comment":"The Top-K/Bottom-K analysis is partially circular. Instances are ranked by BridgeMIL attention weights, which are learned from subject-level labels; selecting the top-K instances and then training an inherited-label classifier on them is likely to select instances that are label-consistent by construction, so the observed accuracy ordering does not independently establish that inherited labels are unreliable. It mainly shows that the attention model's own notion of relevance transfers to a different classifier. Add a non-model-based reliability criterion (e.g., spectral/artifact features or agreement across independently trained instance classifiers), or at least a permutation control, and report paired significance tests for the Top-K vs Bottom-K comparisons.","section":"Section RQ2, Figure 3 right"}],"minor_comments":[{"comment":"The symbol b_a is used for both the input bias inside tanh and the scalar output bias; the dimensions of the two are different. Please use distinct symbols and clarify whether the output gate bias is a scalar or a per-instance quantity.","section":"Eq. (12)"},{"comment":"The right panel does not label its horizontal axis as K (20, 50, 100, 200), and the error bars/confidence regions are not defined. Please add axis labels and a caption statement describing what the shaded bands represent.","section":"Figure 3"},{"comment":"All results are reported as mean±std over five seeds, but no significance tests are provided. Given the small number of subjects, paired bootstrap or permutation tests (or confidence intervals) would make the 'highest mean accuracy in 14 of 15 settings' claim more convincing.","section":"Table 1"},{"comment":"The validation-based checkpoint selection uses different burn-in epochs for different datasets (after epoch 10 for ADFTD, after epoch 15 for Rockhill2021, none for Mumtaz2017). Please explain this choice; otherwise it looks like dataset-specific tuning.","section":"Appendix D"},{"comment":"For the 'w/o Sub-bag' and 'w/o Nearby' ablations, please state explicitly whether the removed loss term is simply dropped or replaced by the remaining term, and whether the loss weights are renormalized. This affects the interpretation of the ablation.","section":"Appendix E, Table 2"},{"comment":"The terms 'seed' and 'fold' are used in overlapping ways (random seed vs. cross-validation fold). Define both once and use them consistently, especially in the scarcity and ablation sections.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about subject-identity leakage is, in my reading, the strongest issue. I do not think it is fatal, but the paper currently lacks the control experiment that would make the central mechanism credible, and the small cohort sizes make external validity a real risk. The empirical comparison is otherwise carefully executed. I would also ask the authors to add a limitations paragraph; the current manuscript contains no explicit discussion of the small-cohort generalization risk."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: BridgeMIL is a legitimate two-stage MIL framework for subject-level EEG diagnosis, and the empirical comparison is the strongest part of the paper. The 14/15 accuracy result checks out arithmetically, the subject-level CV is correct (no subject spans folds), and the ablations are honest. If the gains hold, this is a useful step for EEG-based diagnosis.\n\nWhat's new: the dual-scale Stage 1 alignment—nearby-window pairs plus within-subject sub-bag pairs under VICReg—is a reasonable way to pretrain without inherited instance labels, and the feature-retention term in Stage 2 is a sensible bridge. The evaluation spans three datasets and five backbones, which is more thorough than the typical two-dataset paper. The representation analysis and hyperparameter sensitivity table are a plus.\n\nSoft spots, in rough order of severity:\n\n1. The mechanism is under-tested. The Stage 1 objective makes same-subject windows and sub-bags share representations. That could be disease-relevant, but it could also be encoding subject identity or recording condition. The t-SNE shows distinct subject-wise clusters, and the SC/DB metrics are only computed against diagnostic classes, never against subject identity. The paper does not report a direct test separating these. This is the biggest weakness: the claimed decoupling is plausible but not demonstrated.\n\n2. Hyperparameters (λret, K) are tuned on ADFTD and the ADFTD result is reported with those values. That is a form of selection on the test set for that dataset. Not fatal—the other two datasets were not used for tuning—but the headline ADFTD margin should be read with that in mind.\n\n3. The subject-scarcity analysis changes test-fold composition as the fold count decreases, so the comparison across conditions is confounded. The qualitative trend may still be right, but the numbers aren't clean.\n\n4. No significance tests. Several margins are under two standard errors. The overall 4.28-point gain is more convincing than any single cell.\n\nThe circularity in the Top-K/Bottom-K analysis is minor, since it is an auxiliary illustration.\n\nBottom line: the paper deserves a serious referee. The method is coherent and the empirical scope is above average. The authors should add a subject-identity probe (e.g., linear classification of subject ID from Stage 1 embeddings) and a deconfounded scarcity experiment. I would not reject over the current weaknesses; I would ask for those additions.\n\nI would bring this to reading group and cite it if I were working on EEG or medical time-series MIL.","headline":"Solid method-level paper with a real 14/15 empirical result; the weak spots are tuned hyperparameters and a mechanism that may learn subject identity rather than disease.","tokens_in":19369,"tokens_out":4683,"would_cite":true,"duration_ms":47628,"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":"Decoupling instance learning from subject labels lifts EEG diagnosis accuracy.","keywords":["EEG","disease diagnosis","multiple instance learning","self-supervised representation learning","subject-level supervision","attention MIL","feature retention","medical time series"],"falsifier":"Train Stage 1 with disease labels randomly permuted across subjects while keeping all within-subject structure intact; if subject-level accuracy after Stage 2 remains close to the reported levels, the pretraining signal is not carrying diagnostic information. Conversely, pretrain only on healthy-control recordings and test whether the features transfer to patient classification; poor transfer would indicate the alignments are not disease-relevant.","tokens_in":18343,"feed_emoji":"🧠","tokens_out":5934,"duration_ms":59935,"temperature":0.7,"pith_summary":"The paper claims that the standard practice of training EEG classifiers on short windows with the subject's disease label copied onto every window is a supervision mismatch, because some windows carry little diagnostic evidence. It introduces BridgeMIL, a two-stage framework that first learns instance representations from abundant EEG windows without any inherited labels — by aligning nearby windows and independently sampled sub-bags from the same subject — and then transfers that encoder to an attention-based multiple-instance-learning head that receives supervision only at the subject level. Across three EEG disease datasets and five backbone architectures, BridgeMIL reports the highest mean subject-level accuracy in 14 of 15 settings and an overall mean accuracy of 76.57%, 4.28 percentage points above the strongest baseline. If this holds, it would mean the field's inherited-label training is not just a harmless shortcut but an actual source of error, and that the correct supervision level is attainable without losing access to the abundance of unlabeled EEG instances.","feed_headline":"Two-stage pipeline beats inherited-label EEG training in 14 of 15 settings","feed_subtitle":"Learning features without subject labels, then supervising only the subject prediction, nets 4.28 points.","key_machinery":"The load-bearing mechanism is the two-scale alignment objective in Stage 1 together with the feature-retention bridge into Stage 2. Stage 1 pairs each EEG window with a nearby window and with the mean-pooled representation of an independently sampled sub-bag from the same subject; a variance-invariance-covariance objective (with target standard deviation and off-diagonal covariance penalties) keeps these pairs close without relying on negative pairs, so the encoder learns temporal consistency and within-subject shared structure without needing any instance labels. Stage 2 uses an attention-based MIL aggregator: each instance embedding gets an independent sigmoid gate, the gated embeddings ar","core_discovery":"BridgeMIL's central claim is that subject-level EEG diagnosis should be decomposed into two problems solved at different supervision levels. Stage 1 pretrains the EEG encoder without disease labels using two positive-pair constructions: temporally nearby windows are aligned to capture local consistency, and independently sampled within-subject sub-bags are aligned to capture shared recording structure; variance and covariance regularization prevent the representation from collapsing or becoming redundant. Stage 2 then freezes that representation space as a reference, fine-tunes the encoder inside an attention-based MIL aggregator where a learned gate weights each instance's contribution to t","pith_inferences":["Editorial inference: the same decoupling — unsupervised instance encoding plus labels only at the bag level — should transfer to other per-patient time-series diagnostics, such as arrhythmia detection from ECG, where one label covers a long recording.","Editorial inference: if Stage 1 alignments encode disease-relevant structure, the pretrained encoder should transfer across EEG datasets and recording sites; this could be tested by pretraining on one dataset and fine-tuning the MIL head on another.","Editorial inference: the paper does not directly probe what Stage 1's shared structure represents; linear probes for disease versus subject identity on the pretrained features would settle whether the gain comes from diagnostic content or from subject or artifact alignment.","Editorial inference: the subject-scarcity result implies that collecting more labeled patients should matter more than recording more minutes per patient, a concrete data-collection priority for clinical deployment."],"forward_implications":["If correct, clinical EEG pipelines should stop copying the subject diagnosis onto every window and instead reserve labels for the subject-level aggregation step.","The approach should give its largest gains precisely where labeled subjects are scarce and instances are abundant, since the paper finds subject scarcity is the binding constraint.","Attention weights from the trained model can rank windows by diagnostic relevance, giving a practical way to identify which parts of a recording carry evidence.","The finding that inherited-label reliability varies across instances means instance-level accuracy is not a trustworthy proxy for subject-level diagnosis; evaluation should shift to subject-level metrics."],"fun_headline_variants":["No inherited EEG labels: two-stage model wins 14/15 benchmarks","Pretrain EEG encoder without labels, then supervise subjects only","BridgeMIL separates instance features from subject supervision","4.28-point accuracy gain by decoupling EEG training stages","Rethink EEG diagnosis: learn instances first, supervise subjects later"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes that what nearby EEG windows and separately sampled sub-bags from the same subject share is diagnostic structure; if the dominant shared structure is subject identity, recording-condition drift, or artifacts, the pretrained encoder anchors Stage 2 to non-diagnostic features and the reported gains would not generalize.","fun_headline_variants_meta":{"raw":{"variants":["No inherited EEG labels: two-stage model wins 14/15 benchmarks","Pretrain EEG encoder without labels, then supervise subjects only","BridgeMIL separates instance features from subject supervision","4.28-point accuracy gain by decoupling EEG training stages","Rethink EEG diagnosis: learn instances first, supervise subjects later"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2753,"prompt_tokens":785,"completion_tokens":1968,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":1883}},"tokens_in":529,"tokens_out":1968,"duration_ms":15405,"temperature":1.0,"reasoning_tokens":1883,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T01:39:10.166508+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Stage 1 with disease labels randomly permuted across subjects while keeping all within-subject structure intact; if subject-level accuracy after Stage 2 remains close to the reported levels, the pretraining signal is not carrying diagnostic information. Conversely, pretrain only on healthy-control recordings and test whether the features transfer to patient classification; poor transfer would indicate the alignments are not disease-relevant.","supporting_citations":[],"review_version":2}