{"id":"63087fc2-03a5-4593-a2be-157c0beab1fa","arxiv_id":"1908.10417","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CNN models outperform LSTM, RBM, and wavelet on same-record ECG denoising, but cross-record generalization is poor, with negative SNR on a held-out record.","lead":"A study compares CNN, LSTM, RBM, and wavelet methods for denoising ECG signals, reporting that CNNs perform best when trained and tested on the same patient record. The authors propose training on synthetic 'ECG at effort' data derived from a patient's rest ECG for real-time denoising, but this proposal is not validated on unseen data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated synthetic-to-real transfer: §4.5 tests only on synthetic variants of record 118, never on actual effort ECG; cross-record test (§4.4) already shows negative SNR.","rationale":"The reader's weakest assumption—that synthetic effort ECGs faithfully represent real effort ECGs—is exactly where the central claim hinges, and I agree with it. The paper's positive CNN results in §4.5 are all computed on synthetic test examples generated from the same record and same procedure as training; they do not validate the real-time deployment scenario in the abstract. The single cross-record test in §4.4 is the only distribution-shift evidence and it is negative. A further aggravating factor is that the synthetic dataset is built from a few rest sequences, so random splitting may leak near-duplicate heartbeats across train/test, making even the synthetic generalization numbers optimistic; this follows from the generative procedure described and does not require questioning the authors' intent. The proposed real-effort test would directly settle whether the synthetic-to-real transfer is real. Since the REJECT verdict is already based on this gap, my stress-test does not change it; it sharpens the reason and gives a concrete experiment.","tokens_in":23447,"tokens_out":3279,"duration_ms":35610,"concrete_test":"Use a paired rest/exercise ECG dataset (e.g., MIT-BIH ST Change Database or a recorded exercise stress test). From a subject's rest portion, generate the §4.5 synthetic training set (vary HR 72-90 bpm, add electrode-motion noise at SNR -6, 0, 6, 12), train the CNN with 3 conv layers, 36 filters, and 23x23 kernels, then test on the subject's actual effort segment (with real noise), computing RMS and SNR against the clean or annotated effort ECG. If the average SNR falls below 8 dB or is materially worse than the synthetic held-out 8-10 dB, the transfer claim fails. As a secondary check, compare beat morphology (RR, QT, ST deviation) between synthetic effort beats and true effort beats; systematic differences would identify why the transfer fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Abstract and §4.5 make the central claim that a CNN trained on 1-second synthetic multiple-heartbeat signals, generated by varying heart rate (72-90 bpm) and adding noise to a few rest sequences from an ECG record, can denoise the same patient's ECG at effort. The load-bearing premise is that time-scaled/repeated rest heartbeats plus noise are a faithful proxy for true effort ECG. This premise is never tested. All §4.5 evaluations are on synthetic test signals drawn from the same generator and same record 118 (random 3/4-1/4 split), so they only show interpolation inside the generator's family. Moreover, because every synthetic signal is built from a few rest heartbeats, the random split can place near-duplicate waveforms in training and test, inflating even the synthetic held-out numbers (average SNR 8.2-9.7 dB). The only out-of-family evaluation, §4.4 training on 10 records and testing on record 100, yields average SNR = -2.08 dB, demonstrating failure when the test distribution is not covered by training. Real effort ECG differs from rest ECG beyond heart rate (ST-T changes, QRS/T amplitude changes, exercise-induced baseline wander), so the claimed deployment scenario is unsupported. The paper's own conclusion (§5) describes the real-time use as something that 'would be' useful, not a demonstrated result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares deep learning models (CNN, LSTM, RBM) and conventional/wavelet methods for ECG denoising using synthetic signals and MIT-BIH PhysioNet records. The central proposal is to train a CNN on 1-second synthetic multiple-heartbeat ECG signals generated from a few rest sequences of a given record (record 118), varying heart rate from 72 to 90 bpm and adding noise at SNR levels -6, 0, 6, 12 dB, and then to use the trained CNN to denoise the same patient's ECG at effort (Section 4.5). The paper also reports a 'Design of Experiment' study (Section 4) that selects an optimal CNN architecture, and it compares results on same-record and cross-record settings (Sections 4.1-4.4). The headline same-record results are favorable (e.g., average SNR 15.02 in Section 4.1), but the only held-out-record test in Section 4.4 yields average SNR -2.0776 dB on record 100, and all Section 4.5 evaluations are on synthetic variants of the same record used for training.","tokens_in":23726,"tokens_out":3798,"duration_ms":41877,"significance":"If the central claim were validated, a per-patient CNN trained on rest-derived synthetic ECG data could enable real-time denoising for wearable and exercise ECG monitoring, which would be practically valuable. The paper has useful strengths: it provides detailed network architectures (Tables 1, 2, 5, 6, 9), compares several baselines on the same data, uses publicly available PhysioNet records, and reports the negative cross-record result honestly rather than hiding it. However, the significance is heavily limited because the main deployment claim rests on an untested synthetic-to-real transfer, and the reported performance numbers are not supported by statistical validation. The 'DoE' study, presented as a novelty, is actually a one-factor-at-a-time grid search without the design and analysis machinery usually associated with design of experiments. The paper would need new experiments on real effort ECG data, proper held-out validation across multiple records, and repeated-seed statistics to support its conclusions.","major_comments":[{"comment":"The central claim that a CNN trained on synthetic multiple-heartbeat signals generated from rest ECG can denoise the same patient's ECG at effort is never tested. All Section 4.5 evaluations are on synthetic test signals drawn from the same generator and the same record 118; no real effort ECG signal is used. The load-bearing premise is that time-scaled/repeated rest heartbeats plus added noise faithfully represent effort ECG, but effort ECG differs from rest ECG beyond heart rate (e.g., ST-T changes, amplitude changes, exercise-induced baseline wander). The paper's own Conclusion (Section 5) phrases the real-time use as something that 'would be' useful, not as a demonstrated result. This unsupported transfer is the basis of the abstract's headline claim.","section":"Section 4.5"},{"comment":"The only out-of-family evaluation in the paper, training on 10 records and testing on record 100, gives average RMS 0.3360 and average SNR -2.0776 dB (Section 4.4, final paragraph), which is far below the paper's own 8 dB usefulness threshold stated in Section 2 and far worse than the same-record results (e.g., average SNR 15.02 in Section 4.1). This negative result directly contradicts the generalization needed for the proposed real-time deployment, and it is acknowledged in the text as showing the difficulty of approximating diverse ECG signals with one model. However, the paper then proceeds to Section 4.5 without breaking the self-reference: the proposed per-patient solution is only validated on synthetic variants of the same record, so the cross-record failure remains unresolved.","section":"Section 4.4"},{"comment":"The reported performance metrics lack statistical rigor and are likely optimistic. In Sections 4.1, 4.2, and 4.5, training and testing are done on random splits from the same record, and because the synthetic signals are generated from a small set of rest sequences by time-scaling and repeating heartbeats, the split can place near-duplicate waveforms in both training and test sets, inflating the average RMS/SNR values. Moreover, no error bars, no multiple random seeds, and no repeated cross-validation are reported. The DoE architecture selection in Table 3 also uses the same test split to choose the 'optimal' CNN, so the selected architecture is tuned to that split and the reported SNR values (e.g., 12.36 dB for simulation 46) are not an unbiased estimate of performance.","section":"Sections 4.1 and 4.5, Table 3"},{"comment":"The 'Design of Experiment' study is not a design of experiments in the usual statistical sense. It is a grid search over 51 configurations with different filter counts and kernel sizes, with no factorial structure, replication, blocking, or analysis of variance. The selection rule based on average RMS, average SNR, and computational time is ad hoc, and no statistical significance is attached to the chosen optimum. The abstract and conclusions claim novelty for a DoE study, but the actual methodology is a standard hyperparameter sweep; the claim should be reframed or the methodology upgraded.","section":"Section 4, 'Design of Experiment'"}],"minor_comments":[{"comment":"The abstract mentions 'the standard wavelet-based technique for denoising EEG signals'; this appears to be a typo for ECG signals and should be corrected.","section":"Abstract"},{"comment":"The caption for Figure 23 references 'Fig.22(a)' when describing the second panel; the cross-reference is inconsistent and should be fixed.","section":"Figure 23 caption"},{"comment":"The proposed RMS limit of 0.3 mV (5% of a 6 mV range) is introduced without justification or clinical basis; if it is used as an acceptance criterion, its choice should be explained and its sensitivity assessed.","section":"Section 2"},{"comment":"The exact procedure for generating the multiple-heartbeat synthetic dataset is under-specified: the text says heartbeats are varied between 12 and 15 per 10 seconds, but does not state how rest heartbeats are selected, how time-scaling is applied, or how the electrode motion artefact is scaled to the four SNR levels. This makes reproducibility difficult.","section":"Section 4.5"},{"comment":"The phrase 'first order zero phase delay Butterworth filters' should be 'zero-phase Butterworth filters' for clarity and standard terminology.","section":"Algorithm 1"},{"comment":"Several statements mix 'epochs' and 'iterations' (e.g., Section 4.4 says '250 iterations (i.e. 94400 iterations)'); the terminology should be made consistent and precise.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper's central claim—per-patient real-time ECG denoising from rest-derived synthetic data—is not supported by the experiments as presented: the only held-out-record test fails, and the Section 4.5 evaluation is circular. The DoE novelty claim is also overstated. Should the authors resubmit, they would need to validate on real effort ECG, use properly held-out records and repeated-seed statistics, and provide code or detailed reproducibility materials. The honest reporting of negative cross-record results is a point in the paper's favor, but it does not cure the load-bearing validation gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim is unvalidated. The paper proposes training a CNN on synthetic 1-second multi-heartbeat signals generated from a few rest sequences of a patient, then using it to denoise that patient's ECG at effort. That scenario is never tested. The only cross-record test (Section 4.4) yields negative SNR, and the Section 4.5 tests are synthetic variants of the same record. The paper is honest about this in places—its conclusion says the real-time use \"would be\" useful—but the abstract oversells it.\n\nWhat's worth credit: the paper runs a wide comparison of CNN, LSTM, RBM, and wavelet denoising, with a 51-configuration grid search over filters and kernel sizes. That is a decent amount of work, and the DoE table is a useful reference for how CNN capacity and runtime trade off on 1-second ECG segments. The synthetic augmentation idea is not new, but applying it to patient-specific effort ECG is an interesting direction. The paper also reports run times and some negative results, which is more transparent than many.\n\nSoft spots: the evaluation is circular. The DoE selects an architecture on a test split from the same record, and the headline SNR figures come from the same record. There are no error bars, no multiple seeds, and no code or data, so the numbers are not reproducible. The cross-record result (record 100) is catastrophic: average SNR -2.08 dB. That is not a minor blemish; it directly contradicts the claim that the approach generalizes. The synthetic-to-real transfer premise is also untested: real effort ECG differs from rest ECG in more ways than heart rate (ST-T changes, amplitude changes, baseline wander), and the paper's synthetic generator only varies heart rate and noise. So the deployment scenario is unsupported.\n\nThis is a load-bearing flaw, not a fixable detail. The paper would need a real validation study—train on rest, test on actual effort recordings from the same person, with proper held-out records and statistical rigor—before it merits publication.\n\nWho this is for: someone looking for a negative example of why same-record evaluation is tempting and misleading, or a survey of CNN hyperparameters for ECG denoising. Not for someone who wants a demonstrated method.\n\nRecommendation: I would desk reject this. The idea has some merit, but the current evidence does not justify peer review. Encourage the author to add real effort ECG validation and rerun.","headline":"Patient-specific synthetic ECG augmentation is a reasonable idea, but the paper never validates it: same-record training/testing and a negative cross-record SNR sink the central claim.","tokens_in":24222,"tokens_out":2973,"would_cite":false,"duration_ms":30488,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A CNN trained on synthetic multiple-heartbeat ECG built from a patient's rest recording can, the paper argues, denoise that patient's real effort ECG.","keywords":["ECG denoising","convolutional neural network","synthetic ECG generation","multiple-heartbeat ECG","design of experiments","signal-to-noise ratio","LSTM","Restricted Boltzmann Machine"],"falsifier":"Record one patient's clean resting ECG and then the same patient's ECG during exercise while simultaneously capturing a clean reference from a different lead or a high-grade monitor. Train the paper's CNN on synthetic 1-second effort-like windows generated from the rest ECG, apply it to the real effort recording, and compute SNR and RMS against the clean reference. If the denoised effort ECG does not reach positive SNR or does not preserve R-peak positions, the rest-to-effort transfer premise is falsified.","tokens_in":23251,"feed_emoji":"🫀","tokens_out":8377,"duration_ms":72415,"temperature":0.7,"pith_summary":"This paper argues that a convolutional neural network trained on synthetic one-second ECG segments, generated from a few seconds of a patient's clean resting electrocardiogram by varying heart rate and adding heavy noise, can denoise that same patient's ECG when it is recorded at effort. Across comparisons on real and synthetic ECG data, the CNN outperforms LSTM, RBM, and wavelet denoising in the tested settings, and a Design-of-Experiments study picks a preferred CNN structure with 3 convolutional layers, 96 filters, and 13x1 kernels. The payoff would be personalized, real-time ECG denoising for wearables and ambulatory monitors without requiring large labeled ECG datasets. The paper's cross-record test also shows that a model trained on ten records does not transfer to an eleventh unseen record, where average SNR is negative.","feed_headline":"Synthetic heartbeats can train a CNN to clean effort ECG","feed_subtitle":"A few seconds of rest ECG could generate personalized training data for wearable denoising, the paper argues.","key_machinery":"The central machinery is a 15-layer regression CNN acting on 1-second ECG windows of 360 samples: three convolutional blocks with 36 filters, 23x23 kernels, batch normalization, ReLU, and average pooling with stride [4 1], feeding a fully connected layer and a mean-squared-error regression output. The complementary machinery is the synthetic-data generator: starting from a few real rest heartbeats, it creates multiple-heartbeat 1-second windows at 72-90 beats per minute and adds real electrode-motion plus random noise at SNRs -6, 0, 6, and 12 dB to simulate effort ECG. The DoE sweep over 51 architectures is the selection mechanism: it ranks filter counts from 16 to 96 and kernel sizes from 9x1 to 99x1, plus selected square kernels, by testing RMS, SNR, and runtime, yielding the recommended 96 filters and 13x1 kernels.","core_discovery":"On the paper's own terms, the central discovery is that a regression CNN can learn to map a heavily corrupted 1-second ECG window back to its clean form, and, more specifically, that synthetic multiple-heartbeat training windows generated from a few clean rest sequences of the same ECG record can stand in for real effort ECG. The CNN with three convolutional layers, 36 filters of size 23x23 per layer, batch normalization, ReLU, average pooling with stride 4, and a fully connected regression head reached average testing SNR of 15.02 dB when trained and tested on 1-second windows of one ECG record, and 14.77-16.45 dB on a second record. When trained on synthetic multi-heartbeat data varied over 72-90 beats per minute and noise levels from -6 to 12 dB SNR, it reached average testing SNR of 8.2 dB (or 9.67 dB with a larger training set), and it outperformed LSTM, RBM, and wavelet baselines in the heavy-noise conditions tested. The Design-of-Experiments sweep over 51 filter and kernel combinations identifies 96 filters per layer with 13x1 kernels as the best accuracy-versus-time trade-off for a three-layer CNN.","pith_inferences":["If the rest-to-effort transfer holds, a wearable device could personalize its denoiser in the field: capture a few seconds of clean rest ECG, synthesize an effort-like training set, and fine-tune a small CNN without a central labeled ECG repository.","The negative cross-record result suggests the method's value is personalization rather than a universal denoiser; extending to a new patient would require per-patient synthesis or a domain-adaptation step, neither of which the paper tests.","A decisive same-patient validation would compare rest-trained synthetic denoising on real effort ECG against a simultaneous clean reference; the paper's evaluations stop at synthetic variants of rest data and one negative cross-record test.","The DoE recommendation is tied to one noise type, electrode motion artefact, one sampling rate, and one window length; other noise types or leads could shift the optimal filter and kernel counts."],"forward_implications":["A CNN trained on per-patient synthetic effort-like ECG can denoise multi-heartbeat 1-second windows from the same source, reaching average testing SNR of 8.2 dB with the smaller training set and 9.67 dB with the larger one.","Single-record training and testing reaches average SNR around 15 dB on one record and 14.77-16.45 dB on another, well above the paper's 8 dB threshold for carrying useful information.","The CNN beats the LSTM and RBM baselines in the tested configurations and beats wavelet denoising when the noise is heavy and drifting rather than simply additive.","The DoE sweep shows that accuracy saturates as filters and kernel sizes grow while compute keeps increasing, so the recommended 96-filter, 13x1-kernel architecture is the best available trade-off.","Training on ten records and testing on a different record gives negative average SNR, so a single generic CNN is not sufficient; the paper's personalization route is essential to its proposed use."],"supporting_citations":[{"why":"supplies the dynamical model that generates the synthetic PQRST heartbeat morphology used to build training data","marker":"[23]"},{"why":"provides the real ECG records used for the single-record, multi-record, and cross-record denoising experiments","marker":"[25]"},{"why":"provides the real electrode-motion artefact noise added to clean records at controlled SNR levels","marker":"[26]"},{"why":"supports the heavy-noise test conditions by describing noise levels several times the ECG amplitude","marker":"[24]"},{"why":"is the neural-network ECG denoising baseline whose RBM implementation is compared against the CNN","marker":"[13]"},{"why":"is the LSTM-based ECG denoising baseline that the CNN outperforms in the paper's comparisons","marker":"[16]"},{"why":"defines the pre-filtering algorithm (low-pass, high-pass, notch, baseline wander removal) applied before CNN denoising","marker":"[37]"},{"why":"is the synthetic ECG generation approach the paper adapts for its initial synthetic datasets","marker":"[20]"}],"fun_headline_variants":["Synthetic rest beats train CNN to clean effort ECG","CNN denoises effort ECG after training on synthetic beats","From a few rest beats, CNN learns to denoise effort ECG","Synthetic multi-beat ECG data powers CNN denoising"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that synthetic multiple-heartbeat ECG made from a patient's few rest sequences, by changing heart rate and adding electrode-motion and random noise, is representative enough of that patient's real effort ECG that a CNN trained on the synthetic data will denoise the real signal.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic rest beats train CNN to clean effort ECG","CNN denoises effort ECG after training on synthetic beats","From a few rest beats, CNN learns to denoise effort ECG","Synthetic multi-beat ECG data powers CNN denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1495,"prompt_tokens":1113,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":314}},"tokens_in":729,"tokens_out":382,"duration_ms":4117,"temperature":1.0,"reasoning_tokens":314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:43:47.666812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record one patient's clean resting ECG and then the same patient's ECG during exercise while simultaneously capturing a clean reference from a different lead or a high-grade monitor. Train the paper's CNN on synthetic 1-second effort-like windows generated from the rest ECG, apply it to the real effort recording, and compute SNR and RMS against the clean reference. If the denoised effort ECG does not reach positive SNR or does not preserve R-peak positions, the rest-to-effort transfer premise is falsified.","supporting_citations":[{"cited_title":"The impact of the MIT-BIH Arrhythmia Database","cited_arxiv_id":null,"evidence_quote":"provides the real ECG records used for the single-record, multi-record, and cross-record denoising experiments"},{"cited_title":"A dynamical model for generating synthetic electrocardiogram signals","cited_arxiv_id":null,"evidence_quote":"supplies the dynamical model that generates the synthetic PQRST heartbeat morphology used to build training data"},{"cited_title":"Moody, WE Muldrow, “A noise stress test for arrhythmia detectors“, Computers in Cardiology, 11, 381-384, 1984","cited_arxiv_id":null,"evidence_quote":"provides the real electrode-motion artefact noise added to clean records at controlled SNR levels"},{"cited_title":"Sensors and signal processing methods for a wearable physiological parameters monitoring system","cited_arxiv_id":null,"evidence_quote":"supports the heavy-noise test conditions by describing noise levels several times the ECG amplitude"},{"cited_title":"An ultra low power personalizable wrist worn ECG monitor integrated with IoT infrastructure","cited_arxiv_id":null,"evidence_quote":"defines the pre-filtering algorithm (low-pass, high-pass, notch, baseline wander removal) applied before CNN denoising"},{"cited_title":"Kumar, S","cited_arxiv_id":null,"evidence_quote":"is the synthetic ECG generation approach the paper adapts for its initial synthetic datasets"}],"review_version":1}