{"id":"e5e2cff8-84db-4bf4-a856-584a47eaa841","arxiv_id":"2504.18608","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"An ECG authentication method that pre-trains on signal-text pairs and uses center, prototype, and repulsion losses to reject unregistered identities in open-set tests.","lead":"This paper combines ECG signals with text descriptions of heart waveform features to pre-train an identity recognition model, then adds three training losses to reject strangers not enrolled in the system. It reports 99.83% authentication accuracy and an open-set classification rate above 95%, but the core repulsion loss as written has the wrong gradient direction.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (9) and Algorithm 1 line 15 specify max(d - R, 0), a hinge that pulls known-class features toward their reciprocal points, not away; since the paper's open-set rejection claim rests on this repulsion term, the central mechanism is internally inverted as written.","rationale":"The paper's central claim is open-set rejection via repulsion, and Eq. (9) is the only place this mechanism is defined. Its sign makes it an attraction hinge, which is not a matter of tuning or hyperparameters but of optimization direction. The reader's weakest assumption identified exactly this issue, and the manuscript text in Algorithm 1 line 15 confirms it is not a one-off typo in the prose. This concern is load-bearing regardless of whether the reported tables are reproducible: as specified, the method cannot produce the claimed behavior. The other noted issues, such as the nonstandard FAR definition, the TNR of 49.03% contradicting the claim that most unregistered samples are filtered out, and the recurring Fr/Fs encoder confusion, are serious but secondary. The repulsion-sign error alone is sufficient to reject the central claim. If the authors later release code showing the sign is reversed, the paper could be revised with corrected equations and independent reproduction; until then the central claim is internally inconsistent and unverifiable.","tokens_in":14200,"tokens_out":5272,"duration_ms":56134,"concrete_test":"Re-run the ECGID open-set protocol from Section 3.2 with two variants of Algorithm 1: (i) the loss exactly as printed in Eq. (9), max(d-R,0), and (ii) the sign-reversed repulsion hinge max(R-d,0), keeping all other settings identical. If variant (i) attains OSCR/FAR close to Table 1, the reported gain cannot be attributed to repulsion; if variant (ii) is needed to reproduce the table, Eq. (9) is a typesetting error. Either way, inspect the official code's line for L_o to confirm which form was actually optimized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is 'Irrelevant Sample Repulsion Learning.' Section 2.4 states that reciprocal points P_id should be close to the complementary set D,id_L ∪ D_U (Eq. 8), and that target identity samples should be distanced from these points. The implemented loss in Eq. (9) is L_o = (1/(MN)) Σ max(d_e(F_r(x_i),P_id) - R_id, 0), with the same form repeated in Algorithm 1 line 15. Minimizing max(d - R, 0) reduces d whenever d > R; gradient descent on this term moves samples closer to the reciprocal points. The correct repulsion hinge would be max(R - d, 0), which penalizes only samples that are too close. Thus, if the code follows the printed equation and algorithm, the mechanism advertised as pushing known samples away from irrelevant/open-set features is actually pulling them toward those features. The paper provides no code or hyperparameters to disambiguate this from a typographical error; Algorithm 1 line 15 shows the same wrong sign, so the manuscript as specified does not implement the claimed open-set repulsion. This is load-bearing because the ablation in Table 3 attributes the FAR reduction to component B.1, but that reduction would not be explained by the printed loss.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an ECG identity-authentication system for open-set scenarios. It consists of a multi-modal pretraining stage in which ECG signals are aligned with text reports generated from fiducial features, followed by fine-tuning with three loss components: self-constraint center learning, dynamic prototype learning, and irrelevant sample repulsion learning. The method is evaluated on the ECGID, MIT-BIH, and Autonomic datasets under various open-set ratios, with reported results including a closed-set accuracy of 99.83%, a False Accept Rate as low as 5.39%, and an Open-set Classification Rate above 95%. The authors claim consistent superiority over five baseline methods.","tokens_in":14532,"tokens_out":9742,"duration_ms":100511,"significance":"If implemented correctly, the paper would make a useful contribution to open-set ECG authentication: the evaluation uses held-out identities from three public datasets, the pretraining uses external MIMIC-ECG data, and the ablation study covers the proposed components. The headline numbers are not forced by circularity because the open-set test identities are not used for training. However, the central novelty, Irrelevant Sample Repulsion Learning, is mis-specified in the printed equations and algorithm, and no code or hyperparameters are provided, so the current manuscript does not yet substantiate the claims.","major_comments":[{"comment":"The loss L_o = (1/(MN)) * sum_{id,i} max(d_e(F_r(x_i), P_id) - R_id, 0) is not a repulsion term. When d_e > R_id, gradient descent decreases the loss by decreasing d_e, which moves the sample feature toward P_id; moreover, R_id is itself a learnable parameter updated to minimize the same objective, so the loss is trivially minimized by increasing R_id. The correct repulsion form would be a hinge of the opposite sign, for example max(R - d_e, 0), with R fixed or regularized. As printed, the mechanism advertised as separating registered samples from open-set samples is internally inverted or vacuous, and the experimental results, including the ablation in Table 3, cannot be interpreted as validating the paper's central claim.","section":"Section 2.4, Eq. (9), and Algorithm 1 line 15"},{"comment":"Equation (4) defines C_id as an argmin over m in S_id, but the displayed objective does not reference m, so the formula does not define a center; it should presumably be C_id = argmin_{m in S_id} sum_j d(m, m_j^id). Furthermore, Equations (5), (6), (7), and (9) mix the report encoder F_r with the signal encoder F_s, for example Eq. (5) writes ||F_r(x_i^id) - C_id||_2^2 while Algorithm 1 computes distances on S_k = F_s(x_k). Because of these inconsistencies, the fine-tuning procedure as written cannot be implemented unambiguously.","section":"Section 2.3, Eq. (4), and Eqs. (5)-(9)"},{"comment":"The ablation results do not support the prose attribution of the FAR reduction to the irrelevant-sample repulsion module. Removing B.1 (row with A=yes, B.1=no, B.2=yes, B.3=yes) raises FAR from 7.53% to 8.40%, whereas removing B.2 or B.3 raises FAR to 15.48% and 15.11%, respectively. Combined with the issue in Eq. (9), the claim that B.1 drives the advertised open-set rejection behavior is not established by the reported experiments; the authors need to rerun the ablations with a correctly specified repulsion loss and report which component is actually responsible.","section":"Section 3.5, Table 3"}],"minor_comments":[{"comment":"Equation (3) contains the duplicate term L^{r2s}_{i,j} twice in the summand; it should presumably be L^{s2r}_{i,j} + L^{r2s}_{i,j}.","section":"Section 2.2, Eq. (3)"},{"comment":"The algorithm is titled 'SimCLR's main learning algorithm', but it describes the proposed fine-tuning procedure, not SimCLR; the title should be corrected.","section":"Algorithm 1"},{"comment":"The function zeta(.,.) used in Eq. (8) is never defined, and the connection between the inequality on P_id and the claimed separation of registered samples from open-set samples is not explained.","section":"Equation (8)"},{"comment":"The headline numbers 99.83% accuracy and 5.39% FAR come from different experimental setups: the accuracy is from the Autonomic experiment with 30 enrolled identities, while the FAR comes from the ECGID experiment with 30 enrolled and 11 open-set identities. Reporting them together in the abstract is misleading and should be clarified.","section":"Abstract and Section 3.4"},{"comment":"There are numerous formatting and typographical issues, including 'Hanghzou' in the affiliation, the leftover 'Preprint submitted to Physics Letters B' header, incomplete reference [19], and inconsistent notation for the batch size L versus the loss function L; these should be corrected.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"I am recommending major revision rather than rejection only because the central loss can in principle be repaired by a corrected equation and regularization of the margin. However, the absence of released code makes it impossible to verify whether the reported experiments used the printed loss or a corrected one; if the authors cannot supply the corrected formulation and code, or if the corrected loss changes the reported results, the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: (1) the central open-set repulsion loss is written with the wrong sign—minimizing max(d - R, 0) pulls registered features toward the reciprocal points, not away. (2) The abstract's 5.39% FAR is computed with a nonstandard denominator and sits uneasily with the TNR numbers in Table 1. The paper has a real research idea, but as it stands the mechanism and the metrics don't line up.\n\nThe genuinely new piece is the multimodal pretraining: ECG signals paired with text reports auto-generated from fiducial features (R-peak positions, RR intervals, QRS width, SDNN, RMSSD), then aligned with contrastive learning. That's a plausible way to inject physiological knowledge into the encoder, and it seems to help in the ablations—ACC jumps from 80% to 99.6% on ECGID. The fine-tuning additions (center learning, prototypes, reciprocal points) are adapted from known methods, so the novelty is mostly the assembly.\n\nThe problems are serious. Eq. (9) and Algorithm 1 line 15 both specify max(d - R, 0). Gradient descent on that term decreases d whenever d > R, so samples are pulled toward the reciprocal points—the opposite of the claimed repulsion. The correct hinge would be max(R - d, 0). This is not a small typo: it's the load-bearing mechanism for open-set rejection, and the ablation credits this component for the FAR drop. If the authors' code used the correct sign, the paper needs to say so; if it followed the printed equation, the results are not explained.\n\nThere are also notation problems: Eq. (4) argmin doesn't reference m; Eqs. (5)-(7) use the report encoder Fr for signal inputs; Eq. (3) duplicates one of the contrastive terms. FAR in Eq. (12) divides by |DL|+|DU|, which systematically lowers FAR as the open set grows. And on ECGID the TNR is 49%, so the system accepts over half of unregistered samples—hard to square with 'filter out most unregistered samples' in Section 3.3. No code or hyperparameters are provided, so nothing can be checked.\n\nI'm not convinced the current numbers are trustworthy. The multimodal pretraining direction is worth following up, and if a revision corrects the repulsion loss, cleans up the notation, reports a conventional FAR, and releases code, the paper could become a solid contribution. As is, I would not accept it, but I'd send a revised version back to good reviewers rather than desk-reject the idea outright.","headline":"The paper's load-bearing repulsion loss is inverted as written, and the evaluation metrics don't support the abstract's claims, but the multimodal pretraining idea is worth a second look after major fixes.","tokens_in":15033,"tokens_out":6543,"would_cite":false,"duration_ms":62221,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that open-set ECG identity authentication can be made reliable by aligning ECG signals with fiducial-feature text reports during pretraining and then constraining the feature space with center, prototype, and repulsion…","keywords":["ECG identity authentication","open-set recognition","multi-modal pretraining","contrastive learning","reciprocal points","self-constraint center learning","biometric security"],"falsifier":"Reimplement Eq. (9) exactly as printed and inspect the gradient of the loss with respect to the sample feature: minimizing $\\max(d - R, 0)$ moves the sample toward the reciprocal point, so the advertised repulsion mechanism is inverted. Separately, evaluate the trained model on open-set identities drawn from a dataset not represented among the enrolled classes and check whether the OSCR and FAR remain at the reported levels.","tokens_in":13970,"feed_emoji":"🫀","tokens_out":7688,"duration_ms":65487,"temperature":0.7,"pith_summary":"This paper tries to establish that ECG identity authentication can remain accurate and reject strangers when the test stream contains identities never seen during training. It claims that pretraining the signal encoder with contrastive alignment to text reports of fiducial features, then fine-tuning with three feature-distribution constraints, yields 99.83% authentication accuracy on the closed set and a false accept rate as low as 5.39% when open-set samples are present. Across open-set-to-closed-set ratios up to 1:10, it reports an Open-set Classification Rate above 95%. A sympathetic reader would care because real authentication systems must reject unseen impostors, and most earlier ECG methods were evaluated only with known identities.","feed_headline":"ECG authentication rejects strangers with 5.39% false accepts","feed_subtitle":"Multi-modal pretraining plus repulsion learning keeps OSCR above 95% even at a 1:10 open-set ratio.","key_machinery":"The central machinery is a two-stage training pipeline. Stage one is multi-modal contrastive pretraining: a signal encoder and a text encoder (MedCPT, a pre-trained medical text encoder) are aligned on 100,000 ECG-text pairs drawn from MIMIC-ECG, where the text is a templated report of fiducial features. Stage two is a fine-tuning loss with three terms: $L_{\\text{self}}$ (the L2 distance from each sample to a class center), $L_{\\text{proto}}$ (distance-based classification with learnable prototypes, replacing softmax), and $L_o$ (the irrelevant-sample repulsion term, defined as $\\max(d(\\mathbf{F}(x_i), \\mathbf{P}_{id}) - R_{id}, 0)$ with learnable reciprocal points $\\mathbf{P}$ and margin $R$). Together these losses compress each enrolled identity's features into a compact cluster while separating them from the reciprocal points that represent all other identities; the paper's intended effect is that open-set features, which were never seen, remain in the bounded region outside the registered clusters and are therefore rejected by the decision threshold.","core_discovery":"On the paper's own terms, the central discovery is that open-set ECG identity authentication can be made accurate without ever showing the model a true open-set sample during training. The method first aligns ECG waveforms with automatically generated text reports of five fiducial features (R-peak positions, RR intervals, QRS width, SDNN, RMSSD) using contrastive pretraining on a large signal-text dataset. During fine-tuning, three losses reshape the feature space: a self-constraint center loss pulls each identity's samples toward a class center, a dynamic-prototype distance loss replaces softmax classification with learnable prototypes, and an irrelevant-sample repulsion loss, inspired by adversarial reciprocal points, is intended to push registered samples away from the regions where other identities (serving as pseudo-open-set samples) lie. The paper reports this yields 99.83% closed-set authentication accuracy, a false accept rate as low as 5.39% when open-set samples are present, and an OSCR above 95% across open-set ratios up to 1:10.","pith_inferences":["The reported OSCR was measured on open-set identities drawn from the same source populations as the enrolled classes; if real-world impostors occupy feature regions the enrolled classes never covered, the true open-set rejection rate could be lower than reported.","The reciprocal-point repulsion idea is not specific to ECG; the same fine-tuning recipe could be tested on other physiological signals or on face and fingerprint authentication, where open-set rejection is also a practical need.","A falsifiable prediction follows from the method's mechanism: registered-sample features should be farther from the reciprocal points than open-set features are. That distance gap can be measured directly on any of the three datasets and compared with the claimed FAR."],"forward_implications":["If the central claim holds, a deployed ECG authentication terminal can keep enrolled-user accuracy near 99.8% even when the input stream contains many people who are not registered.","The reported FAR of 5.39% means that in a small open-set setting, roughly 19 out of 20 unregistered users are rejected rather than mistaken for an enrolled identity.","Stability across open-set ratios up to 1:10 means the method does not need to know in advance how many impostor identities will appear.","Ablations show all three losses contribute: dropping any component raises FAR or lowers OSCR, so the gains are not attributable to pretraining alone.","The method transfers across three datasets with different sampling rates (500, 360, and 1000 Hz), suggesting the pipeline is not tuned to one acquisition device."],"supporting_citations":[{"why":"Supplies the 800,035-pair MIMIC-ECG signal-text dataset from which 100,000 pairs are used for multi-modal pretraining.","marker":"[7]"},{"why":"Provides the MedCPT text encoder that encodes the fiducial-feature reports in the contrastive pretraining step.","marker":"[13]"},{"why":"Introduces adversarial reciprocal points learning, the idea that Irrelevant Sample Repulsion Learning adapts for open-set separation.","marker":"[5]"},{"why":"The MERL multimodal ECG-text method that motivates using clinical text alongside ECG and serves as a baseline to beat.","marker":"[18]"},{"why":"The Zhou multi-hypersphere open-world ECG method, the strongest baseline compared in the open-set experiments.","marker":"[34]"},{"why":"Supplies the ECGID dataset used for fine-tuning and for the small-scale open-set evaluation where FAR 5.39% is reported.","marker":"[19]"},{"why":"Supplies the MIT-BIH Arrhythmia database used as a second fine-tuning and open-set evaluation dataset.","marker":"[20]"},{"why":"Supplies the Autonomic Aging dataset used for the large-scale open-set ratio experiments and the 99.83% accuracy report.","marker":"[26]"}],"fun_headline_variants":["Open-set ECG auth: 99.83% accuracy, 5.39% FAR","ECG auth repels strangers with self-constraint learning","Multimodal ECG pretraining tames open-set impostors","ECG auth keeps 95% OSCR with repulsion learning","Self-constraint centers + repulsion: ECG auth for unknowns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the repulsion loss as implemented actually pushes registered samples away from the reciprocal points—the printed equation would pull them closer—and that samples from other enrolled identities faithfully stand in for never-seen impostors.","fun_headline_variants_meta":{"raw":{"variants":["Open-set ECG auth: 99.83% accuracy, 5.39% FAR","ECG auth repels strangers with self-constraint learning","Multimodal ECG pretraining tames open-set impostors","ECG auth keeps 95% OSCR with repulsion learning","Self-constraint centers + repulsion: ECG auth for unknowns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001063,"raw_usage":{"total_tokens":4480,"prompt_tokens":990,"completion_tokens":3490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3396}},"tokens_in":606,"tokens_out":3490,"duration_ms":25632,"temperature":1.0,"reasoning_tokens":3396,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:20:14.654529+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reimplement Eq. (9) exactly as printed and inspect the gradient of the loss with respect to the sample feature: minimizing $\\max(d - R, 0)$ moves the sample toward the reciprocal point, so the advertised repulsion mechanism is inverted. Separately, evaluate the trained model on open-set identities drawn from a dataset not represented among the enrolled classes and check whether the OSCR and FAR remain at the reported levels.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 800,035-pair MIMIC-ECG signal-text dataset from which 100,000 pairs are used for multi-modal pretraining."},{"cited_title":"Medcpt: Contrastive pre-trained transformers with large- scale pubmed search logs for zero-shot biomedical information retrieval","cited_arxiv_id":null,"evidence_quote":"Provides the MedCPT text encoder that encodes the fiducial-feature reports in the contrastive pretraining step."},{"cited_title":"Adversarial reciprocal points learning for open set recognition","cited_arxiv_id":null,"evidence_quote":"Introduces adversarial reciprocal points learning, the idea that Irrelevant Sample Repulsion Learning adapts for open-set separation."},{"cited_title":"Open-world electrocardiogram classification via domain knowledge- driven contrastive learning","cited_arxiv_id":null,"evidence_quote":"The Zhou multi-hypersphere open-world ECG method, the strongest baseline compared in the open-set experiments."},{"cited_title":"Biometric human identification based on electro- cardiogram","cited_arxiv_id":null,"evidence_quote":"Supplies the ECGID dataset used for fine-tuning and for the small-scale open-set evaluation where FAR 5.39% is reported."},{"cited_title":"The impact of the mit-bih arrhythmia database","cited_arxiv_id":null,"evidence_quote":"Supplies the MIT-BIH Arrhythmia database used as a second fine-tuning and open-set evaluation dataset."},{"cited_title":"Autonomic aging–a dataset to quantify changes of cardiovascular autonomic function during healthy aging","cited_arxiv_id":null,"evidence_quote":"Supplies the Autonomic Aging dataset used for the large-scale open-set ratio experiments and the 99.83% accuracy report."}],"review_version":1}