{"id":"210f8fe9-1c7a-40a3-943d-9f99316cd50a","arxiv_id":"2511.16627","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"TFCDiff, a diffusion model trained on truncated DCT coefficients of 10-second ECG segments with time-frequency feature fusion, outperforms eight benchmark denoisers, including on the unseen real SimEMG noise dataset.","lead":"TFCDiff cleans noisy heart recordings by running a diffusion model on the frequency-domain (DCT) coefficients of 10-second ECGs, with a time-frequency feature-fusion module that recovers waveform details. It reports better denoising than eight existing methods on simulated noise and, more strikingly, on a real muscle-artifact dataset the model never saw during training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DCT truncation (50-Hz cutoff) gives TFCDiff an untested band-limiting advantage; SimEMG superiority may stem from hard low-pass filtering rather than the learned denoiser.","rationale":"I chose the bandwidth confound over the reader's ground-truth concern because it directly affects the comparative SOTA claim. The ground-truth preprocessing issue applies to all methods and would primarily affect absolute performance, not the relative ranking; the band-limiting issue gives TFCDiff a structural advantage that the baseline models do not have, so the reported superiority on SimEMG could vanish if baselines are given the same cutoff. The paper already includes a public code link and a reasonably controlled synthesized benchmark, but no control that isolates the truncation. A simple post-hoc filtering of baseline outputs or a truncation-only baseline would resolve the ambiguity. Therefore the verdict remains CONDITIONAL, with the additional condition that the authors run this control. I disagree with the reader's identification of the weakest assumption because their focus on preprocessing fidelity is less likely to change the relative assessment; the bandwidth inequality is more fixable and more decisive.","tokens_in":20560,"tokens_out":8392,"duration_ms":75942,"concrete_test":"On the SimEMG test set (Section IV-B), apply a 50-Hz zero-phase low-pass filter to the denoised outputs of TCDAE and DesCod-10 (or equivalently, DCT-truncate their outputs with the same 1000-coefficient cutoff) and recompute the five metrics of Table IX. Also compute a truncation-only baseline by DCT-truncating the noisy input, zero-filling above 50 Hz, and applying IDCT. If either TCDAE+filter or the truncation-only baseline reaches TFCDiff-1's SSD/ImSNR, the SOTA generalization claim is largely attributable to band-limiting rather than the proposed diffusion/TFEM architecture.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing untested assumption is the fairness of the bandwidth comparison. In Section III-A, TFCDiff truncates DCT coefficients to the first 1000 (for N=3600 at fs=360 Hz), i.e., all signals are hard band-limited to 50 Hz. Consequently, any high-frequency EMG/MA noise above 50 Hz is removed before the model sees it, and the output (after zero-padding and IDCT) contains no energy above 50 Hz. All eight baselines (TCDAE, DesCod, DeepFilter, etc.) operate on full-band time-domain signals and must learn to suppress these components. On the SimEMG database, which contains real broadband EMG noise, this gives TFCDiff a prior-based advantage that is independent of the diffusion process or TFEM. The paper's ablation (Table X) compares DCT-only vs. time-domain diffusion on the synthesized set, but does not test a simple truncation-only denoiser or apply the same band-limit to baselines. Thus the central generalization claim (TFCDiff-3 beats all benchmarks on all metrics) is confounded: the improvement may reflect the fixed 50-Hz cutoff rather than the learned conditional denoiser. The paper's own limitation (Section VI-D: 'unsatisfactory denoising results for signals with intensive EMG noise') further suggests that the method's apparent robustness to EMG on SimEMG may be an artifact of the truncated bandwidth.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TFCDiff, a conditional diffusion model for denoising 10-second ambulatory ECG signals. The method operates in the DCT domain: clean and noisy signals are transformed by DCT, truncated to the first 1000 coefficients (≈50 Hz at fs=360 Hz), and used as diffusion training targets and conditioning inputs. The noise predictor is a 1D U-Net augmented with a Temporal Feature Enhancement Mechanism (TFEM) that interleaves time-domain and frequency-domain feature streams. Training data are QT Database segments preprocessed with bandpass/median filtering and piecewise-linear baseline removal, corrupted by flexible random mixtures of BW, MA, and EM noise from the NST Database. The model is evaluated on a synthesized QT/NST test set and on the external SimEMG database, with comparisons against eight baselines. Table III reports that TFCDiff-10 achieves the best mean values on all five metrics for the synthesized set, while Table IX shows TFCDiff-3 and TFCDiff-1 outperforming all benchmarks on the SimEMG set. The paper also includes ablations for the DCT-domain design and TFEM, plus generation-time measurements.","tokens_in":20872,"tokens_out":5266,"duration_ms":48451,"significance":"If the results hold, TFCDiff addresses a real gap: multi-beat, raw-sequence ECG denoising without beat segmentation, which is more clinically relevant than single-beat evaluation. The external SimEMG validation is a genuine strength, as is the public code release and the systematic ablation of the time-frequency design. The DCT-domain diffusion idea with TFEM is technically interesting and the reported computational costs are plausible. However, the central SOTA and generalization claims currently rest on a comparison that may be confounded by the hard 50-Hz band-limit inherent to the DCT truncation, and the abstract overstates the synthesized-set result by omitting the ensemble-size dependence. The underlying method is defensible and the issues are addressable with additional control experiments, so the contribution is potentially significant for wearable ECG denoising.","major_comments":[{"comment":"The DCT truncation to 50 Hz gives TFCDiff a structural advantage that is not controlled for. All inputs and outputs are hard band-limited by zeroing DCT coefficients above [50/Δf] (≈1000 of 3600 coefficients at fs=360 Hz). In contrast, all eight baselines operate on full-band time-domain signals. On SimEMG, which contains broadband real EMG noise, simply truncating the DCT of the noisy input and applying IDCT may remove a large fraction of the noise without any learned denoiser. The paper does not include a truncation-only baseline, nor does it band-limit the baselines. Consequently, the claim in Table IX that TFCDiff-3 outperforms all benchmarks on all metrics is confounded: the improvement could be largely due to the fixed 50-Hz cutoff rather than the learned conditional denoiser or TFEM. Please add a control that applies the same 50-Hz truncation to the noisy input (and optionally to","section":"§III-A, §V-B, Table IX"},{"comment":"The state-of-the-art claim in the abstract ('achieves state-of-the-art performance across five evaluation metrics') is unqualified. Table III shows that TFCDiff-1 is worse than TCDAE on SSD, PRD, and CosSim; TFCDiff-3 lags on SSD, PRD, and CosSim; and only TFCDiff-10 leads on all five metrics. The ensemble size k is a critical hyperparameter that determines the headline result. Please qualify the SOTA claim with the ensemble size, and either state TFCDiff-10 explicitly or report the trade-off between k and performance. The conclusion also repeats the unqualified claim.","section":"Abstract, §V-A, Table III"},{"comment":"The ground-truth x0 used for training and for evaluating the synthesized set is not the raw QT recording but a surrogate: bandpass/median-filtered, then piecewise-linear baseline-removed within QRS-annotated segments, with Hermite interpolation at junctions. The paper does not validate this surrogate against the original signal or an independent reference. If residual baseline wander, filter ringing, or interpolation artifacts remain, the model is trained to reproduce those artifacts and the reported SSD/MAD/PRD/CosSim/ImSNR measure reconstruction to the processed surrogate, not to the true clean ECG. Please provide a quantitative or visual validation of the preprocessing (e.g., comparison of the surrogate with the raw signal in low-noise segments, or an expert review), and discuss how any residual artifacts affect the interpretation of the intra-dataset results.","section":"§IV-B"},{"comment":"The ablation in Table X does not isolate the effect of the 50-Hz truncation from the effect of DCT-domain modeling. The standalone DCT-diffusion row (SSD 146.8, CosSim 0.727) is much worse than the time-domain baseline (SSD 59.5, CosSim 0.889), and adding TFEM reverses this. This suggests that TFEM is compensating for information lost by truncation or by the skewed DCT-coefficient scaling, but the reader cannot tell how much of the final advantage comes from the hard band-limit versus from the learned denoiser. Please include a control where a time-domain model is trained on the same truncated-DCT-reconstructed signals (i.e., band-limited inputs without the DCT diffusion objective), or equivalently, where the time-domain baselines receive a 50-Hz low-passed version of the input.","section":"§VI-A, Table X"}],"minor_comments":[{"comment":"Duplicate sentence: 'Both transforms must be orthogonal to ensure invertibility.' appears twice consecutively.","section":"§III-A"},{"comment":"The name 'DesCod' is inconsistently written as 'DesCod' and 'DeScoD' (e.g., Section IV-D vs. Section V-A and reference [18]). Please unify.","section":"Throughout"},{"comment":"CBAM-DAE reports PRD = 1910.425±877.430%, which is physically implausible for ECG denoising; consider explaining this outlier or omitting it from the table with a note, as is already done for later tables.","section":"Table III"},{"comment":"The text says all methods use input length 3600, with exceptions for FCN-DAE and CBAM-DAE at 3584. Please clarify how the outputs are aligned to the 3600-sample evaluation window (padding? cropping?) so that the metric comparison is unambiguous.","section":"§IV-D"},{"comment":"The caption of Fig. 4(b) does not identify which curve corresponds to which SNR scaling factor c. Please add a legend.","section":"Fig. 4"},{"comment":"The PRD definition in Eq. (19) uses the denoised signal's mean in the denominator; this differs from some prior definitions that normalize by the clean signal's mean. Please ensure this is intentional and state it clearly, since PRD values are otherwise hard to compare across papers.","section":"§IV-C"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the band-limiting confound. I would not reject the paper, because the external SimEMG result is genuinely interesting and the method is well-motivated, but the SOTA and generalization claims are not established without a truncation-only control or band-limited baselines. Please ensure the revision addresses this with a concrete experiment. The abstract's unqualified SOTA claim should also be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a serious, well-executed application of DCT-domain conditional diffusion to multi-beat ECG denoising, with an external test on real EMG-contaminated data that is genuinely informative. Second, the strongest generalization claim is partly confounded by a hard 50-Hz band limit that the baselines don't get, and the paper doesn't test whether truncation alone is doing the work.\n\nWhat's actually new: training the diffusion model on truncated DCT coefficients of 10-second ECG segments, the TFEM cross-domain feature mixing, and an SNR rescaling (c=150) that handles the skewed DCT spectrum. The core ideas come from image-domain papers (DCTdiff, cross-field fusion), but the 1D ECG-specific adaptation is new, and the paper is honest about that. The SimEMG results are the strongest part: on an unseen real-world dataset, TFCDiff-3 beats all eight benchmarks on all five metrics. That is real evidence of transfer, and the code is available.\n\nThe soft spots are in proportion. The biggest is the bandwidth confound: TFCDiff truncates DCT coefficients to 50 Hz before inference, so its output contains no energy above 50 Hz. Baselines operate on full-band signals and have to learn to suppress broadband EMG. On SimEMG, which is full of broadband EMG, this is a serious advantage independent of the learned denoiser. The paper's own limitation section admits poor results on intensive EMG noise, which fits the truncation-artifact story. The fix is straightforward conceptually: add a truncation-only baseline and apply the same band limit to the baselines. Without that, the SimEMG superiority claim is conditional, not established.\n\nOther issues are minor but worth fixing: no significance testing; the PRD denominator is nonstandard (it uses the denoised signal against the mean of the ground truth), and the abstract's SOTA claim is unqualified even though the single-generation model is not SOTA on the synthesized set. The ablation table is helpful and honest—DCT-only is worse than time-domain, and TFEM is what makes the combination work.\n\nBottom line: this is a competent paper with a genuine external validation and an addressable confound. It deserves a serious referee, not a desk reject, and I'd want the revision to include the band-limit ablation and significance tests. If the band-limit issue is resolved, it's a useful contribution to wearable ECG denoising.","headline":"Serious, well-executed DCT-domain diffusion for multi-beat ECG denoising with a genuine external test, but the SimEMG advantage is likely inflated by an untested 50-Hz band limit.","tokens_in":21429,"tokens_out":2767,"would_cite":true,"duration_ms":26885,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TFCDiff claims that ECG denoising works best when a conditional diffusion model runs in the DCT domain on raw 10-second multi-beat recordings, with a temporal-feature enhancement stream, rather than operating beat-by-beat in the time domain","keywords":["ECG denoising","conditional diffusion model","discrete cosine transform","time-frequency fusion","multi-beat ECG","random mixed noise","wearable monitoring","generalization"],"falsifier":"Take raw multi-beat ECG with a simultaneously recorded reference clean lead (e.g., shoulder electrodes), apply TFCDiff without any preprocessing, and measure residual error with and without applying the paper's own bandpass/median/piecewise-baseline pipeline to the reference. If the model's advantage over the strongest non-diffusion baseline shrinks or reverses when the reference is not pipeline-processed, the claimed gain is at least partly an artifact of learning the preprocessing. Additionally, injecting a noise type absent from training (e.g., 50-Hz powerline interference) and observing wh","tokens_in":20398,"feed_emoji":"🫀","tokens_out":6286,"duration_ms":58157,"temperature":0.7,"pith_summary":"The paper tries to establish that ECG denoising of long, multi-beat recordings improves when the denoiser is a conditional diffusion model operating on truncated discrete-cosine-transform (DCT) coefficients of the noisy signal, reinforced by a temporal-feature stream inside the noise-prediction U-Net. A sympathetic reader would care because previous deep denoisers mostly handle single segmented heartbeats, while wearable monitors produce continuous 10-second records with mixed baseline wander, muscle artifact, and electrode motion noise. The authors report that on a synthesized multi-beat test set, their model with ten averaged generations scores best among eight compared methods on all five metrics (e.g., SSD 21.5 au², ImSNR 14.1 dB), and that with only one generation it beats the strongest non-diffusion baseline on a real, unseen EMG-contaminated ECG database. They interpret this as evidence that frequency-domain diffusion reconstructs physiologically faithful waveforms, especially T-waves, and generalizes to noise distributions not seen in training. The central claim is conditional: if the preprocessed ground truth and synthetic noise mixtures are accepted, DCT-domain diffusion with time-frequency fusion is the best recipe for wearable ECG denoising.","feed_headline":"DCT-domain diffusion beats eight ECG denoisers","feed_subtitle":"Trained on raw 10-second multi-beat recordings, it also wins on an unseen real-noise ECG dataset.","key_machinery":"The central machinery is conditional diffusion in the DCT domain. A 10-second ECG is DCT-transformed, truncated to the first ~1000 coefficients (below 50 Hz), and both clean target and noisy conditioning input are scaled by a percentile-derived bound (η≈3) from the DC component. The noise predictor is a 1-D U-Net with a Temporal Feature Enhancement Mechanism: Temporal Feature Extraction converts frequency maps to time via zero-padding and IDCT, processes them, and returns via DCT; Temporal Feature Fusion applies time-domain attention at the middle encoder layer. A detour-resample strategy converts maps to time before down/upsampling to avoid aliasing, and the SNR schedule is rescaled by fact","core_discovery":"The paper's finding is that a conditional diffusion model trained and sampled entirely in the truncated DCT domain—conditioned on the noisy signal's DCT coefficients, with a U-Net that intermittently converts feature maps to the time domain—outperforms time-domain diffusion, autoencoders, and filters on raw 10-second ECG segments. The authors report best results with an SNR-scaling factor of 150, a DC-derived scaling bound, and k-generation averaging. On the synthesized test set, TFCDiff-10 achieves SSD 21.5 au² and ImSNR 14.1 dB, surpassing all eight benchmarks; on an unseen real EMG-contaminated database, even one-generation TFCDiff beats the strongest non-diffusion baseline on all five me","pith_inferences":["If the DCT-domain advantage generalizes, the same truncated-DCT conditioning plus time-frequency fusion could apply to other quasi-periodic biosignals (PPG, EEG, EMG) where morphology preservation is diagnostically valuable.","The claimed generalization to real EMG noise is tested only against EMG-type interference; testing against real electrode-motion or baseline-wander recordings would clarify whether the model's robustness extends beyond the noise type present in the unseen test set.","The ground-truth preprocessing dependency suggests a discriminating experiment: compare TFCDiff on raw multi-beat signals against a simultaneously recorded clean lead without the paper's filtering/baseline pipeline; if the margin narrows, the model may be learning the preprocessing rather than true ECG structure.","The authors note unsatisfactory results for intense EMG noise and suggest flow matching; a direct measure of residual noise power inside the QRS band as a function of input SNR would quantify this ceiling and guide whether a different generative formulation is needed."],"forward_implications":["A single model can denoise raw 10-second multi-beat ECG without beat segmentation or padding, removing a barrier to wearable deployment.","Training under flexible random mixed noise (random weights for baseline wander, muscle artifact, and electrode motion, intensity λ∈[0.2,2]) yields robustness across noise levels up to the hardest 1.5–2.0 intensity bin.","With one generation, TFCDiff already outperforms the strongest non-diffusion baseline on an unseen real EMG-contaminated database, suggesting cross-dataset generalization rather than overfitting to synthesized noise.","Averaging more generations (3, 5, 10) monotonically improves all metrics, so users can trade compute (0.17–1.63 s per 10-s clip) for fidelity.","Denoised waveforms visually preserve T-wave morphology that time-domain diffusion smooths away, which matters for clinical interpretation of repolarization abnormalities."],"fun_headline_variants":["DCT diffusion cleans 10-second ECGs, beats 8 denoisers","Time-frequency diffusion denoises ECG, wins on unseen real data","Diffusion in DCT domain tops eight ECG denoising methods","Robust ECG denoising via DCT diffusion beats all benchmarks","Plug-and-play DCT diffusion for ECG noise, SOTA on real data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the filtered, baseline-corrected reference ECG recordings used as ground truth are faithful clean signals, and that amplitude-normalized mixtures of three archived noise excerpts at intensity λ∈[0.2,2] span real ambulatory noise.","fun_headline_variants_meta":{"raw":{"variants":["DCT diffusion cleans 10-second ECGs, beats 8 denoisers","Time-frequency diffusion denoises ECG, wins on unseen real data","Diffusion in DCT domain tops eight ECG denoising methods","Robust ECG denoising via DCT diffusion beats all benchmarks","Plug-and-play DCT diffusion for ECG noise, SOTA on real data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1409,"prompt_tokens":765,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":509,"tokens_out":644,"duration_ms":7054,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T21:06:47.480597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take raw multi-beat ECG with a simultaneously recorded reference clean lead (e.g., shoulder electrodes), apply TFCDiff without any preprocessing, and measure residual error with and without applying the paper's own bandpass/median/piecewise-baseline pipeline to the reference. If the model's advantage over the strongest non-diffusion baseline shrinks or reverses when the reference is not pipeline-processed, the claimed gain is at least partly an artifact of learning the preprocessing. Additionally, injecting a noise type absent from training (e.g., 50-Hz powerline interference) and observing wh","supporting_citations":[],"review_version":1}