{"id":"b6e4a1bf-05be-4556-81e9-b3d07d1433ce","arxiv_id":"2508.20398","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A TransUNet-style 1D network with a hybrid time-frequency loss achieves MAE 0.1285 and PCC 0.9540 on mixed-noise ECG denoising at 0 dB.","lead":"This paper builds a transformer-based U-Net that denoises ECG signals using a loss combining time-domain and frequency-domain accuracy. The authors report better error and correlation metrics than three baseline networks on MIT-BIH data corrupted with NSTDB noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"'Consistent superiority' rests on a single 0 dB comparison with no error bars or strong baselines; statistical and baseline checks are needed.","rationale":"The reader's weakest_assumption points to the synthetic-noise proxy as the core risk. I agree that external validity is a concern, but the more immediate and technically load-bearing issue is internal evidential support: the paper's flagship claim of 'consistent superiority' is backed only by a single-condition comparison with three baselines and no uncertainty quantification. This is not a proof of falsehood, but it means the central claim is currently underdetermined. The proposed concrete test directly targets whether the observed margins are robust and whether the claim generalizes across conditions and stronger baselines. Since the reader's verdict was already CONDITIONAL, my read does not change the verdict; it sharpens the specific conditions that must be met. I mark agreement as partial because the reader focused on the synthetic-noise proxy, while I focus on the statistical and baseline-adequacy gap, though both fall under the umbrella of insufficient evaluation evidence.","tokens_in":6250,"tokens_out":6134,"duration_ms":71184,"concrete_test":"Run the identical train/test protocol five times with different seeds for TF-TransUNet1D and all baselines, and report mean±std for MAE, PCC, SNRI. Apply a paired Wilcoxon signed-rank test (or bootstrap) on the same test segments for TF-TransUNet1D vs. the best baseline. If the improvement in MAE is not significant at p<0.05, the 'consistent superiority' claim is unsupported. In addition, extend Table 1 to include at least two non-learning baselines (wavelet thresholding, EMD) and two recent deep baselines (e.g., Descod-ECG, a Transformer-based denoiser) across all SNR conditions (0/5/10 dB) and individual noise types; report per-condition win counts. If TF-TransUNet1D is not best on a majority of conditions, 'consistent superiority' should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that TF-TransUNet1D shows 'consistent superiority' over state-of-the-art baselines in ECG denoising. The only head-to-head table (Table 1) compares the model to just three baselines (CNN-LSTM, U-Net 1D, FastRNN) under a single condition: mixed noise at 0 dB. No standard deviations, confidence intervals, repeated runs, or significance tests are reported. The broader results in Fig. 2, which cover multiple noise types and SNRs, show no baseline lines, so they cannot substantiate 'consistent' superiority. Because the reported 24.3% MAE improvement over U-Net 1D could be within training-run variance or specific to the 0 dB composite-noise mixture, the headline claim is not yet established. This is a load-bearing evidential gap: if the margin disappears under repeated runs or under stronger recent baselines (e.g., diffusion-based Descod-ECG, wavelet/EMD denoisers), the paper's central claim fails. Additionally, the training section notes that all models were trained for 100 epochs and share hyperparameters, but no model selection or early-stopping details are given, so the comparison may not be fair across models.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes TF-TransUNet1D, a 1D U-Net with a Transformer encoder at the bottleneck and a hybrid time-frequency loss for ECG denoising. The model is trained and evaluated on synthetic noisy segments from the MIT-BIH Arrhythmia Database and the NSTDB. The authors compare TF-TransUNet1D with CNN-LSTM, U-Net 1D, and FastRNN under mixed noise at 0 dB, reporting best MAE (0.1285), PCC (0.9540), and SNRI (13.36). The paper also reports qualitative waveform results and claims robust generalization across noise types and SNR levels.","tokens_in":6610,"tokens_out":4612,"duration_ms":49245,"significance":"If substantiated, the proposed architecture is a plausible combination of established components, and the dual-domain loss is a reasonable mechanism for spectral preservation. The manuscript clearly describes the model and the synthetic evaluation protocol, and it reports standard denoising metrics. However, the empirical evidence is currently too narrow to support the headline 'consistent superiority' claim: no repeated runs, confidence intervals, or significance tests are reported, the baseline set is very small, and several recent methods cited in the paper are not included in the comparison. The contribution is therefore incremental and needs stronger validation before publication.","major_comments":[{"comment":"The central claim of 'consistent superiority' is supported by a single head-to-head table at one condition (mixed noise, 0 dB) with only three baselines. No standard deviations, confidence intervals, or significance tests are reported, so the 24.3% MAE improvement over U-Net 1D may be within run-to-run variance. Figure 2 reports TF-TransUNet1D only; without baseline lines it cannot establish consistent superiority across noise types and SNRs. Please add repeated runs with uncertainty measures and report baseline curves in Fig. 2.","section":"§3.2, Table 1"},{"comment":"The paper does not report the loss weights w_time and w_spectral, the Smooth L1 threshold beta, or the Transformer depth, number of heads, and embedding dimension used in the experiments. It also states all models were trained for 100 epochs with 'same settings' but gives no early-stopping or model-selection criterion. Without these numbers, the comparison is not fully reproducible and the fairness of cross-model comparison is unclear. Please report full hyperparameters and the selection protocol.","section":"§3.1 Implementation and §2.3 (Eq. 5)"},{"comment":"The paper cites Descod-ECG (ref. 12), wavelet (ref. 3), EMD (ref. 7), and other recent denoising methods as part of the motivation, but none of these appear in the quantitative comparison. For a claim of 'state-of-the-art baselines', the comparison set is too limited. Please include at least one recent deep-learning denoiser and one classical spectral/wavelet baseline, or temper the claim to the three implemented baselines.","section":"§1 and Table 1"},{"comment":"The evaluation relies solely on synthetic noise from NSTDB mixed with MIT-BIH recordings at fixed SNRs, z-normalization, and 10-second sliding windows. While this is a standard stress test, the paper's stated application is ambulatory monitoring and digital twins, where real noise statistics and signal correlations differ. The robustness/generalization claim needs either a real-noise or cross-dataset evaluation, or an explicit limitation paragraph. As written, the external validity is unverified.","section":"§3.1 Noisy Signal Synthesis and Preprocessing"}],"minor_comments":[{"comment":"Equation (1) is referenced but not displayed; please number and show all equations, including Eqs. (2) and (3) if used.","section":"§2.2"},{"comment":"The text reports specific MAE/PCC/SNRI values from Fig. 2, but the figure caption does not describe axis labels or markers, and the figure is not visible enough in the text to verify the claimed numbers. Please ensure every value quoted in the text can be read directly from the figure.","section":"§3.2 and Fig. 2"},{"comment":"'Burst Noise (BW)' appears inconsistent with the standard abbreviation BW for baseline wander. Also, the exact composition of the 'mixed noise' in Table 1 is not specified (which noise types and at what proportions).","section":"§3.2"},{"comment":"The PRD metric is defined but never reported. Either report PRD in the results or remove it from the evaluation metrics section.","section":"§3.1"},{"comment":"The conclusion describes the design as 'lightweight', but no parameter count, FLOPs, or inference time is provided. Please quantify the computational cost if the real-time deployment claim is to be supported.","section":"§4"},{"comment":"Minor typo: 'NVIDIA GeForce RTX 4090GPU' should have a space before 'GPU'. Please proofread the manuscript for similar spacing and capitalization issues.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the proposed method is reasonable, but the empirical evaluation needs substantial strengthening before the central claims can be accepted. The authors should be encouraged to provide code/data, full hyperparameters, repeated-run statistics, and a stronger baseline set. I see no circularity problem; the issue is purely evidential."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is an engineering combination, not a new architectural principle: a 1D U-Net with a Transformer bottleneck plus a time-frequency loss. Second, the headline claim of “consistent superiority” is much broader than the evidence. If you need a preprocessing baseline for ECG denoising, this paper is worth a look; if you want to trust the numbers, wait for code and repeated runs.\n\nWhat it does well: the construction is sensible, the data is standard PhysioNet (MIT-BIH plus NSTDB), and the test-set evaluation on held-out subjects is genuine, not circular. The dual-domain loss is a reasonable way to keep spectral content, and the three baseline comparisons show real improvements on this particular test set. It is honest progress within an established program, no more.\n\nThe soft spots are load-bearing. Table 1 compares only three baselines under one condition (mixed noise at 0 dB), with no standard deviations, no repeated seeds, and no significance tests. The newer SOTA they cite, Descod-ECG, is absent from the comparison. Figure 2 shows their model alone, no baseline curves, so it cannot substantiate “consistent” superiority. Training details are thin: loss weights, Smooth-L1 beta, transformer depth/heads/dim, and parameter counts are missing. Sharing hyperparameters across all models is not a guarantee of fairness without model selection details. These gaps matter because a 24.3% MAE improvement over U-Net 1D could shrink or vanish under proper repeats or stronger baselines. The core idea does not hinge on that exact margin, but the paper's main claim does.\n\nMinor: the synthetic noise proxy is a standard shortcut, but it limits claims about real ambulatory or digital-twin noise. The digital-twin motivation is framing, not validation—there is no downstream task tested. And the “lightweight” claim has no evidence behind it.\n\nWho is this for? A researcher looking for a reasonable Transformer-UNet baseline in ECG denoising, or someone building a preprocessing stack who wants a candidate model to benchmark against. It is not a breakthrough, but it is a coherent piece of applied work.\n\nMy recommendation: send it to peer review with the expectation of major revision. A serious referee should ask for more baselines, uncertainty quantification, full hyperparameters, and code. Desk rejection would miss a usable if under-evidenced contribution.","headline":"Plausible engineering combo of Transformer U-Net and spectral loss; claims of consistent superiority outrun the single-condition, no-error-bar evidence.","tokens_in":6986,"tokens_out":1751,"would_cite":false,"duration_ms":24100,"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":"This paper claims TF-TransUNet1D, a U-Net with a Transformer bottleneck and a time-frequency hybrid loss, consistently outperforms three ECG denoising baselines under mixed noise, reaching MAE 0.1285 and PCC 0.9540 at 0 dB.","keywords":["ECG denoising","TransUNet","time-frequency loss","spectral fidelity","cardiac digital twin","self-attention","signal-to-noise ratio improvement"],"falsifier":"Run the same model on ECG plus noise recorded from real wearable motion artifacts (e.g., collected during walking) rather than NSTDB mixtures. If the SNRI gain over the second-best baseline is no longer significant—say, within a few tenths of a dB—the claim of consistent robustness under realistic conditions is falsified. Alternatively, measure the 60 Hz powerline residual: if a time-domain-only loss already suppresses it, the spectral-loss mechanism is not the driver.","tokens_in":6206,"feed_emoji":"🫀","tokens_out":6644,"duration_ms":63293,"temperature":0.7,"pith_summary":"The paper tries to establish a single architecture for ECG denoising that does not trade off local waveform fidelity against global context. It argues that a U-Net encoder-decoder with a Transformer at the bottleneck captures both, and that a loss combining smooth-L1 time-domain error with FFT magnitude error preserves spectral content. The authors show on synthetic mixtures of MIT-BIH ECG with NSTDB noise that this combination, TF-TransUNet1D, achieves the best MAE, PCC, and SNR improvement among four models tested at 0 dB. If true, the result matters because cardiac digital twins depend on artifact-free ECG inputs to maintain waveform features like P-waves and ST-segments in personalized simulations. The paper positions the model as a lightweight plug-in for real-time monitoring pipelines.","feed_headline":"ECG denoiser hits 0.1285 error, topping three baselines","feed_subtitle":"Hybrid loss keeps waveform shape and spectrum intact, targeting reliable inputs for cardiac digital twins.","key_machinery":"The carrying mechanism is TF-TransUNet1D itself: a 1D U-Net with four down/up-sampling levels, a standard Transformer encoder (multi-head self-attention plus feedforward, with residual connections and layer normalization) inserted at the deepest bottleneck, and skip connections that concatenate encoder features to the decoder. The input is a 3,600-sample (10-second) ECG segment; the encoder compresses it to a short, high-channel feature sequence, the Transformer models long-range dependencies across it, and the decoder up-samples back to a denoised waveform. The training objective is a weighted sum of a time-domain Smooth L1 loss and a spectral loss comparing FFT magnitudes of the clean and","core_discovery":"On the paper's own terms, the central discovery is that adding a Transformer encoder at the bottleneck of a 1D U-Net and training with a weighted dual-domain loss—smooth L1 in time plus L2 on FFT magnitudes—yields consistent superiority over CNN-LSTM, 1D U-Net, and FastRNN baselines. At a 0 dB mixed-noise target, the model reaches a mean absolute error of 0.1285, a Pearson correlation of 0.9540, and an SNR improvement of 13.36 dB, reducing MAE by about 24.3% relative to the second-best model. The authors interpret this as evidence that the spectral loss prevents the network from over-smoothing, preserving diagnostic waveform components such as distinct P-waves, sharp QRS complexes, and T-wav","pith_inferences":["If the synthetic-noise assumption holds only loosely, the reported 0.9540 PCC may degrade on real ambulatory recordings; a natural test is evaluating on noise recorded from wearable sensors rather than NSTDB mixtures.","The paper does not report statistical significance or variance over seeds; the 24.3% MAE advantage could shrink on re-runs, so the claim of consistent superiority would be stronger with confidence intervals.","A direct ablation—removing only the spectral term—would isolate how much of the 13.36 dB SNRI gain comes from the frequency loss versus the Transformer bottleneck; the architecture suggests the two contribute additively but this is not shown.","The same dual-domain loss could transfer to other one-dimensional physiological signals such as EEG or fetal ECG, where spectral fidelity is similarly diagnostic, though the paper does not test this."],"forward_implications":["Denoised outputs can feed downstream cardiac digital twin simulations with higher fidelity, since P-wave and ST-segment morphology is claimed to survive denoising.","At 0 dB mixed noise, SNRI reaches 13.36 dB and MAE drops 24.3% versus the next best model in the comparison, so the method delivers a large margin at the hardest tested SNR.","The lightweight design is claimed to make real-time ambulatory monitoring and wearable deployment feasible, a direct corollary of the architecture's modest overhead.","The same Transformer-in-U-Net recipe with dual-domain loss is put forward as a reusable strategy for other biomedical time-series, not just ECG.","The frequency loss should make the model less prone to over-smoothing, which is the paper's explanation for its edge under heavy noise."],"supporting_citations":[{"why":"Motivates cardiac digital twins that need high-fidelity ECG inputs, establishing why denoising quality matters.","marker":"[1]"},{"why":"Further digital-twin context, grounding the claim that waveform alterations cascade into modeling errors.","marker":"[2]"},{"why":"Represents conventional wavelet-transform denoising that the paper argues distorts morphology.","marker":"[3]"},{"why":"Fully convolutional denoising autoencoders, an autoencoder baseline family the paper's architecture builds on.","marker":"[9]"},{"why":"Deep recurrent neural network for ECG denoising, the RNN baseline family (likely FastRNN).","marker":"[10]"},{"why":"Attention-based denoising autoencoder, a comparison point for attention mechanisms in ECG denoising.","marker":"[13]"},{"why":"Disentangled autoencoder method, another autoencoder baseline family referenced alongside [9] and [11].","marker":"[16]"},{"why":"Used to support the claim that networks often overlook spectral reconstruction, motivating the spectral loss component.","marker":"[18]"}],"fun_headline_variants":["Transformer U-Net denoises ECG with hybrid loss, beats 3 baselines","ECG denoising via TF-TransUNet1D: dual-domain loss, 0.1285 MAE","Hybrid time-frequency loss boosts ECG denoising for digital twins","TF-TransUNet1D: spectral loss preserves P-waves, QRS, T-waves","SNR up 13.36 dB: Transformer U-Net cleans ECG for digital twins"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper's numbers assume that noise synthesized by adding NSTDB recordings to MIT-BIH ECG at fixed SNRs, after z-normalization and 10-second windowing, behaves like real-world noise in ambulatory and digital-twin settings; if real noise has different statistics or correlates with the signal, the reported error and correlation metrics will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Transformer U-Net denoises ECG with hybrid loss, beats 3 baselines","ECG denoising via TF-TransUNet1D: dual-domain loss, 0.1285 MAE","Hybrid time-frequency loss boosts ECG denoising for digital twins","TF-TransUNet1D: spectral loss preserves P-waves, QRS, T-waves","SNR up 13.36 dB: Transformer U-Net cleans ECG for digital twins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1217,"prompt_tokens":828,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":280}},"tokens_in":572,"tokens_out":389,"duration_ms":4518,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:03:58.636783+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same model on ECG plus noise recorded from real wearable motion artifacts (e.g., collected during walking) rather than NSTDB mixtures. If the SNRI gain over the second-best baseline is no longer significant—say, within a few tenths of a dB—the claim of consistent robustness under realistic conditions is falsified. Alternatively, measure the 60 Hz powerline residual: if a time-domain-only loss already suppresses it, the spectral-loss mechanism is not the driver.","supporting_citations":[{"cited_title":"Medical Image Analysis 101, 103472 (2025)","cited_arxiv_id":null,"evidence_quote":"Motivates cardiac digital twins that need high-fidelity ECG inputs, establishing why denoising quality matters."},{"cited_title":"IEEE Transactions on Medical Imaging 43(7), 2466–2478 (2024)","cited_arxiv_id":null,"evidence_quote":"Further digital-twin context, grounding the claim that waveform alterations cascade into modeling errors."},{"cited_title":"ISA Transactions 114, 251–262 (2021)","cited_arxiv_id":null,"evidence_quote":"Represents conventional wavelet-transform denoising that the paper argues distorts morphology."},{"cited_title":"-Y.: Noise reduc- tion in ECG signals using fully convolutional denoising autoencoders","cited_arxiv_id":null,"evidence_quote":"Fully convolutional denoising autoencoders, an autoencoder baseline family the paper's architecture builds on."},{"cited_title":"Deep Recurrent Neural Networks for ECG Signal Denoising","cited_arxiv_id":"1807.11551","evidence_quote":"Deep recurrent neural network for ECG denoising, the RNN baseline family (likely FastRNN)."},{"cited_title":"IEEE Transactions on Instrumentation and Measurement 71, 1–10 (2022)","cited_arxiv_id":null,"evidence_quote":"Attention-based denoising autoencoder, a comparison point for attention mechanisms in ECG denoising."},{"cited_title":"Electronics 12(7), 1606 (2023)","cited_arxiv_id":null,"evidence_quote":"Disentangled autoencoder method, another autoencoder baseline family referenced alongside [9] and [11]."},{"cited_title":"Biomedical Signal Processing and Control 50, 62–71 (2019)","cited_arxiv_id":null,"evidence_quote":"Used to support the claim that networks often overlook spectral reconstruction, motivating the spectral loss component."}],"review_version":1}