{"id":"29496d41-3fa3-44db-b1e6-27e581b4ba95","arxiv_id":"2501.18367","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised framework that adds AE-GAN reconstruction error from external healthy data to multi-view contrastive learning improves low-label EEG and ECG disease classification in reported experiments.","lead":"The paper proposes LMCRD, a self-supervised framework that adds an AE-GAN reconstruction error feature from external healthy data to a multi-view contrastive encoder for medical time series. It reports consistent accuracy gains over seven baselines on EEG and ECG datasets used for Alzheimer's, Parkinson's, and myocardial infarction diagnosis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AE-GAN's reconstruction-error feature—claimed to encode disease probability—is never validated against cross-center shift; the one ablation that would test it is corrupted by table duplication, so the SOTA claim rests on an untested mechanism.","rationale":"The paper's contribution is not merely a new contrastive loss; it is the claim that AE-GAN reconstruction discrepancy transfers prior knowledge from external healthy data and 'reconstructs target data discrepancies as disease probabilities' (abstract). If that scalar feature is not actually ordered by disease, then the method reduces to LMCF plus an uninformative or confounded channel, and the claimed advantage over COMET/CLOCS is not attributable to the proposed mechanism. The manuscript asserts the ordering in Section 4.1.3 without a single experiment measuring it. The ablation that could partially substitute for such a test (Table B3) is corrupted by copy-paste errors: its FFT 100% LMCRD row equals the PFT row (80.48) rather than the Table 2 FFT value (93.23), and the row named LMCRD-0 (84.50) exceeds the full model's 80.48. With Tables 2 and 3 being exact duplicates, the numerical record is too unreliable to support the strongest claim. These issues are not about disagreement with prevailing consensus; they are about the internal support for the central claim. They can be settled by a targeted diagnostic (reconstruction-error discriminability) plus code release, so they do not force a rejection, but they do make acceptance conditional on evidence that is currently missing. This aligns with the reader's CONDITIONAL verdict, so I recommend UNCHANGED.","tokens_in":17765,"tokens_out":7589,"duration_ms":76751,"concrete_test":"Pretrain the AE-GAN exactly as described on the external healthy set, then compute E on held-out target healthy and target diseased samples from each of the three datasets. Report the AUROC of E alone for disease classification, and also compare E between external healthy and target healthy. If AUROC is near chance or the external-vs-target-healthy difference is comparable to the healthy-vs-disease difference, the transfer mechanism fails. In the same code release, regenerate Table B3's FFT 100% row; it must match the corrected Table 2 value rather than the duplicated PFT value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1.3 (Eq. 2) assumes that because the AE-GAN is pretrained on healthy external samples, the reconstruction error E on target samples is smaller for target-healthy than for target-diseased data. For this to hold, the external-to-target healthy distribution shift must be smaller than the healthy-to-disease signal in the target center; otherwise E encodes center identity, and the appended feature (Eq. 3) injects a center confound into every downstream representation. The paper supplies no evidence for this ordering: no E-by-group summary, no AUROC of E alone, no comparison of E on external healthy vs. target healthy, and no sensitivity analysis. The one ablation intended to show that the AE-GAN module contributes, Table B3, is internally inconsistent: the FFT 100% LMCRD row repeats the PFT value 80.48 instead of Table 2's 93.23, and the 'LMCRD-0' row shows 84.50, i.e., the ablated model appears to beat the full model. Tables 2 and 3 are identical duplications. Thus the central claim that LMCRD consistently outperforms seven baselines, and specifically that AE-GAN knowledge transfer is responsible for part of the gain, is not supported by trustworthy experimental evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LMCRD, a two-stage framework for medical time-series diagnosis. In the first stage, an AE-GAN is pretrained on normal samples from an external dataset and used to compute a reconstruction error for target samples, which is appended as an extra feature. In the second stage, a learnable multi-view contrastive framework (LMCF) with multi-head attention is trained using subject-wise, trial-wise, epoch-wise, temporal-wise, inter-view, and intra-view contrastive losses, followed by supervised fine-tuning. Experiments on an Alzheimer's disease EEG dataset, a Parkinson's disease EEG dataset (TDBrain), and a myocardial infarction ECG dataset (PTB) are reported against seven baselines under partial fine-tuning, full fine-tuning with 100% labels, and full fine-tuning with 10% labels.","tokens_in":18038,"tokens_out":5251,"duration_ms":52507,"significance":"If the results were validated, the paper would address a practically important problem: learning from label-limited, single-center medical time series with the help of external normal data. The proposed combination of an external-data-based reconstruction discrepancy and a multi-head attention contrastive framework is a reasonable direction, and the reported gains are large. However, the experimental support in the current manuscript is not trustworthy as presented: Tables 2 and 3 are identical duplicates, the ablation table contains internal contradictions, and the AE-GAN reconstruction-error mechanism that is central to the claimed advantage is never directly validated. These issues make it impossible to assess whether the reported state-of-the-art performance is real or a consequence of the extra feature and reporting inconsistencies.","major_comments":[{"comment":"Tables 2 and 3 are identical duplicates. The paper claims two full fine-tuning results on the AD dataset, but the same table is printed twice. As a result, the full fine-tuning evidence for the central claim that LMCRD consistently outperforms baselines is currently a single table, not two independent sets of results. The authors must identify which table is the correct one, provide verified results for the other, and explain how this duplication occurred.","section":"§5.2.2, Tables 2 and 3"},{"comment":"Table B3 is internally inconsistent and cannot support the stated ablation conclusions. In the FFT 100% block, the LMCRD row shows accuracy 80.48, which is the PFT value from Table 1, not the 93.23 reported in Table 2. In the same block, the LMCRD-0 row shows accuracy 84.50, which would mean the ablated model outperforms the full model. The text in Appendix E states that LMCRD-0 has the lowest accuracy of 77.42 and that the drops for LMCRD-LMC, LMCRD-RD, and LMCRD-0 are 3.06%, 0.88%, and 3.06%, respectively, which does not match the numbers in Table B3. The ablation therefore does not support the claim that the AE-GAN and LMCF components are indispensable.","section":"Appendix E, Table B3"},{"comment":"The reconstruction-error feature is never validated, yet it is a load-bearing part of the method. The paper asserts that because the AE-GAN is pretrained on external healthy samples, the reconstruction error E is smaller for target healthy samples and larger for abnormal target samples. This requires the external-to-target healthy distribution shift to be smaller than the healthy-to-disease signal in the target center. No evidence is supplied: there are no E-by-group statistics, no AUROC of E alone, no comparison of E on external healthy versus target healthy samples, and no sensitivity analysis. Without such evidence, E may encode center identity rather than disease, and appending E to every target sample (Eq. (3)) could inject a center confound into all downstream representations. In addition, baselines receive F features while LMCRD receives F+1 features, so the comparison is confounded by the extra input channel. The authors should provide direct validation of E as a disease indicator and an ablation that controls for the feature-dimension increase.","section":"§4.1.3, Eqs. (2)-(3)"},{"comment":"The contrastive losses as written do not follow the InfoNCE form that the paper claims. In each of these equations, the denominator sums over negative samples only and does not include the positive term from the numerator. For example, Eq. (5) has numerator exp(h1_i · h2_i) but the denominator sums only over j ≠ i, and Eq. (8) similarly omits the positive pair from the partition function. This is a meaningful technical discrepancy: the optimization objective is not standard InfoNCE, and the behavior of the method may differ substantially from the stated design. The authors should correct the formulas to match their implementation and verify whether the implementation also omits the positive terms.","section":"§4.2, Eqs. (4), (5), (7), (8), (9)"},{"comment":"The claim that LMCRD consistently outperforms seven baselines in both PFT and FFT settings is not supported by trustworthy experimental evidence as presented. Beyond the duplicated main tables and contradictory ablation, there is no significance test or confidence-interval analysis for the differences between LMCRD and COMET, which is the closest baseline. Given the large standard deviations reported in several rows, e.g., LMCRD accuracy 93.23±5.25 in Table 2, the authors should report paired statistical tests or effect sizes to establish that the improvements are not within noise.","section":"§5.2, Tables 1-3"}],"minor_comments":[{"comment":"The method is referred to as both LMCF and LMCRD in the abstract and introduction; the relationship between the framework name and the overall method name should be clarified and used consistently.","section":"Abstract and §1"},{"comment":"The notation in Eq. (9) is under-specified: the meanings of v+, s-, and the summation index j are not clearly defined, and the expression g(x^k_{i,s}) does not show the view argument v even though the output is written as g^k_{i,v,s}.","section":"§4.2, Eq. (9)"},{"comment":"The main text mentions both subject-dependent and subject-independent splits for the external AD dataset, while the target dataset description states a patient-independent split; please clarify which split is used for the reported numbers.","section":"Appendix C"},{"comment":"The paper states that source code is released at 'xxxxx'; a working repository URL or an explicit statement of availability is needed.","section":"General"},{"comment":"The Subject-wise Contrastive Loss paragraph is duplicated verbatim in the text; remove the duplicate.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The duplication of Tables 2 and 3 and the internal contradictions in Table B3 go beyond ordinary presentation slips; they undermine the empirical foundation of the paper. If the authors resubmit a revision, I would ask for the original experimental logs and code to verify that the reported numbers correspond to actual runs. Acceptance is not possible without a full re-review of corrected and re-executed experiments, including the missing validation of the reconstruction-error mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper has a plausible new idea: take an AE-GAN trained on external healthy data, compute per-sample reconstruction error, and append that as an extra channel before contrastive pretraining. Second, the experiments as printed do not support the claimed SOTA. Tables 2 and 3 are identical duplicates; the ablation table contradicts itself; and the central mechanism is never validated.\n\nWhat is actually new is narrow but real. The hierarchical losses come straight from COMET, and the MHA-based views come from LNT. The novelty is the AE-GAN reconstruction-discrepancy feature (Eq. 2-3) and its injection into the contrastive loss. That's a new combination, and the paper is honest that the rest is inherited. The 10%-label results, if real, would be practically useful for medical time-series classification.\n\nThe writing is clear, the method is well-specified, and the hierarchy discussion is correct. But the experimental section is the load-bearing wall, and it has cracks. Table 2 and Table 3 are exactly the same, which is either a copy-paste error or a sign of sloppy reporting. Table B3 is worse: the FFT 100% LMCRD row shows 80.48, which is the PFT number, not the 93.23 in Table 2. And the LMCRD-0 row shows 84.50, meaning the ablated model appears to beat the full model. The text says LMCRD-0 accuracy is 77.42, which matches neither row. That's not a minor typo; it undermines the ablation logic.\n\nThere is also a confound. LMCRD gets the reconstruction-error feature; none of the baselines do. So even the clean-looking comparisons in Table 1 are comparing apples to oranges. The authors could fix this by adding baselines with the extra feature, or by removing it from LMCRD in a controlled way.\n\nThe bigger conceptual gap is the claim that E is a disease probability. That requires the external-to-target healthy distribution shift to be smaller than the healthy-to-disease signal. The paper offers no evidence: no E by group, no AUROC of E alone, no sensitivity to which external center is used. Without that, E might encode center identity, and the method could be injecting a confound into every representation.\n\nThis paper deserves a serious referee, but only if the editor is willing to send it back for major revision. The idea is worth testing, but the current version should not be accepted or cited. I'd want the authors to release code, fix the tables, add the controlled comparison, and provide a direct analysis of E across healthy/disease and across centers. If they do that, it could be a useful contribution for the SSL-for-medical-time-series crowd.","headline":"Plausible new feature, but the experiments as printed are unreliable: duplicate tables, a contradictory ablation, and an unvalidated reconstruction-error mechanism.","tokens_in":18600,"tokens_out":3450,"would_cite":false,"duration_ms":32395,"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":"The paper proposes that reconstruction error from an AE-GAN trained on healthy external data can serve as a disease-probability prior, and that a learnable multi-view contrastive framework built on this prior outperforms seven baselines…","keywords":["medical time series","contrastive learning","self-supervised learning","AE-GAN","multi-view learning","EEG","ECG","disease diagnosis"],"falsifier":"Take the external healthy dataset and the target dataset from two different centers, train the AE-GAN on the external healthy samples, and plot the reconstruction-error distributions for target healthy versus target diseased samples; if the two distributions are not well separated, or if healthy samples from a third center produce errors as large as the diseased target samples, then the discrepancy is encoding center identity rather than disease. A second check is to shuffle the reconstruction-error feature across target samples or replace it with errors from an untrained AE-GAN and observe whether the reported accuracy drop disappears, which would show the feature itself is not the cause of the gains.","tokens_in":17561,"feed_emoji":"🧠","tokens_out":8874,"duration_ms":77608,"temperature":0.7,"pith_summary":"The paper proposes that medical time-series diagnosis can be improved by borrowing healthy-signal knowledge from an external dataset and by replacing hand-crafted contrastive pairs with learnable views. It introduces LMCRD, which first trains an AE-GAN on normal patients from another center, then uses the AE-GAN's reconstruction error on each target sample as an added disease-probability feature, and finally trains a multi-head attention encoder with inter-view and intra-view contrastive losses across subject, trial, epoch, and temporal levels. The claim is that this combined pipeline consistently outperforms seven baselines on Alzheimer's disease, myocardial infarction, and Parkinson's disease datasets, including when only 10% of labels are available. A reader should care because label scarcity and cross-center differences are the main practical obstacles to deploying deep learning on EEG and ECG data.","feed_headline":"Healthy-signal reconstruction error sharpens medical time-series diagnosis","feed_subtitle":"A contrastive framework turns external normal recordings into disease-probability features and beats seven baselines.","key_machinery":"The load-bearing mechanism is the reconstruction-discrepancy feature: a generator trained on healthy external samples produces small error $E = \\mathrm{MSE}(G(\\bar{x}_i), \\bar{x}_i)$ for healthy target samples and larger error for diseased ones, and that scalar is concatenated to the raw sequence so the contrastive encoder can use abnormality likelihood as an extra channel. On top of that, the LMCF encoder combines a dilated convolutional network $h(x)$ with a multi-head attention network $g(x)$ that creates $V$ learnable views; these views are trained by inter-view loss (different views of the same sample are negative) and intra-view loss (same view and subject are positive, different subjects negative), alongside subject-, trial-, epoch-, and temporal-wise InfoNCE losses, weighted $1:1:1:1:2$. The multi-head attention is what removes the need to hand-define positive and negative pairs.","core_discovery":"On its own terms, the paper claims that a single framework can transfer healthy priors across medical centers and adaptively build multi-view representations without manually designing positive and negative pairs. The central assertion is that reconstruction error from an AE-GAN trained only on external healthy samples acts as a usable indicator of disease probability in target data, and that integrating this discrepancy into a learnable multi-view contrastive objective yields state-of-the-art accuracy, precision, recall, F1, AUROC, and AUPRC on Alzheimer's EEG, myocardial-infarction ECG, and Parkinson's EEG benchmarks. The paper further claims that the gains persist and even strengthen in an extreme 10% labeled-data regime, and its ablations attribute the improvement to both the AE-GAN knowledge-transfer module and the multi-view contrastive module.","pith_inferences":["A natural extension is to test the same reconstruction-discrepancy prior on other label-scarce physiological signals, such as sleep staging, seizure detection, or ICU monitoring, where an external healthy cohort is available.","The method's success likely depends on the external and target datasets sharing recording hardware and montage; applying it to a target center with very different equipment may require re-estimating the AE-GAN on matched healthy subjects.","One could isolate the source of gain more sharply by comparing against an AE-GAN trained on target healthy samples only, which would reveal whether the benefit comes from external transfer or from any accurate normal-distribution model.","Because the view contrastive losses are label-free, the framework could be stacked on top of other pretraining objectives, potentially turning it into a general-purpose medical time-series encoder."],"forward_implications":["On all three diseases, the full LMCRD model beats TS2Vec, TF-C, Mixing-up, TS-TCC, SimCLR, CLOCS, and COMET across the reported metrics.","With only 10% of labels for fine-tuning, LMCRD still leads, for example 94.67% accuracy on AD versus 91.43% for the best baseline COMET.","Ablations show that removing the AE-GAN knowledge transfer (LMCRD-LMC) or the multi-view contrastive module (LMCRD-RD) lowers accuracy and F1, so both components contribute.","The method uses the full subject-trial-epoch-temporal hierarchy plus an additional view level, unlike baselines that use only a subset of these levels.","The reconstruction-discrepancy feature is produced without any labels on the target set, so the transfer step needs only unlabeled target data."],"supporting_citations":[{"why":"Supplies the autoencoder-with-GAN architecture used to learn the healthy-signal distribution.","marker":"[13]"},{"why":"Supports using a GAN-based autoencoder to detect anomalies in normal data, the transfer premise.","marker":"[27]"},{"why":"Furnishes the four-level hierarchical contrastive formulation that LMCF extends and is also the strongest baseline.","marker":"[26]"},{"why":"Motivates learning view representations through local transformations, implemented here with multi-head attention.","marker":"[22]"},{"why":"Defines InfoNCE, the loss underlying all the paper's contrastive objectives.","marker":"[21]"},{"why":"Baseline TS2Vec, which contrasts instances and timestamps and is outperformed by LMCRD.","marker":"[28]"},{"why":"Target Alzheimer's EEG dataset used for downstream diagnosis evaluation.","marker":"[9]"},{"why":"External Alzheimer's EEG dataset whose healthy subjects pretrain the AE-GAN.","marker":"[20]"},{"why":"External PTB-XL ECG dataset used to supply normal samples for the myocardial infarction task.","marker":"[24]"},{"why":"TDBrain target dataset used for Parkinson's disease evaluation.","marker":"[23]"}],"fun_headline_variants":["Learnable contrastive views plus reconstruction error boost medical diagnosis","AE-GAN reconstruction error becomes disease probability in contrastive learning","Multi-view contrastive learning with AE-GAN priors beats seven baselines","Reconstruction discrepancy from healthy signals sharpens time-series diagnosis","Adaptive multi-view contrastive learning with reconstruction priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an AE-GAN trained on healthy people at an external center will reconstruct healthy target patients well and sick target patients poorly, meaning the cross-center recording differences are smaller than the disease-related differences in the signal.","fun_headline_variants_meta":{"raw":{"variants":["Learnable contrastive views plus reconstruction error boost medical diagnosis","AE-GAN reconstruction error becomes disease probability in contrastive learning","Multi-view contrastive learning with AE-GAN priors beats seven baselines","Reconstruction discrepancy from healthy signals sharpens time-series diagnosis","Adaptive multi-view contrastive learning with reconstruction priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2634,"prompt_tokens":917,"completion_tokens":1717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1631}},"tokens_in":533,"tokens_out":1717,"duration_ms":13339,"temperature":1.0,"reasoning_tokens":1631,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:44:01.914804+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the external healthy dataset and the target dataset from two different centers, train the AE-GAN on the external healthy samples, and plot the reconstruction-error distributions for target healthy versus target diseased samples; if the two distributions are not well separated, or if healthy samples from a third center produce errors as large as the diseased target samples, then the discrepancy is encoding center identity rather than disease. A second check is to shuffle the reconstruction-error feature across target samples or replace it with errors from an untrained AE-GAN and observe whether the reported accuracy drop disappears, which would show the feature itself is not the cause of the gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the autoencoder-with-GAN architecture used to learn the healthy-signal distribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports using a GAN-based autoencoder to detect anomalies in normal data, the transfer premise."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Furnishes the four-level hierarchical contrastive formulation that LMCF extends and is also the strongest baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline TS2Vec, which contrasts instances and timestamps and is outperformed by LMCRD."},{"cited_title":"Escudero, D","cited_arxiv_id":null,"evidence_quote":"Target Alzheimer's EEG dataset used for downstream diagnosis evaluation."},{"cited_title":"Miltiadous, K","cited_arxiv_id":null,"evidence_quote":"External Alzheimer's EEG dataset whose healthy subjects pretrain the AE-GAN."},{"cited_title":"Wagner, N","cited_arxiv_id":null,"evidence_quote":"External PTB-XL ECG dataset used to supply normal samples for the myocardial infarction task."},{"cited_title":"Van Dijk, G","cited_arxiv_id":null,"evidence_quote":"TDBrain target dataset used for Parkinson's disease evaluation."}],"review_version":1}