Pith. sign in

REVIEW 4 major objections 6 minor 52 references

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

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 21:06 UTC pith:G2GQP33E

load-bearing objection 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. the 4 major comments →

arxiv 2511.16627 v2 pith:G2GQP33E submitted 2025-11-20 eess.SP

TFCDiff: Robust ECG Denoising via Time-Frequency Complementary Diffusion

Pengxin Li , Yimin Zhou , Jie Min , Yirong Wang , Wei Liang , Qingling Xia , Wang Li This is my paper
classification eess.SP
keywords ECG denoisingconditional diffusion modeldiscrete cosine transformtime-frequency fusionmulti-beat ECGrandom mixed noisewearable monitoringgeneralization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

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.

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 (4)
  1. [§III-A, §V-B, Table IX] 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
  2. [Abstract, §V-A, Table III] 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.
  3. [§IV-B] 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.
  4. [§VI-A, Table X] 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.
minor comments (6)
  1. [§III-A] Duplicate sentence: 'Both transforms must be orthogonal to ensure invertibility.' appears twice consecutively.
  2. [Throughout] The name 'DesCod' is inconsistently written as 'DesCod' and 'DeScoD' (e.g., Section IV-D vs. Section V-A and reference [18]). Please unify.
  3. [Table III] 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.
  4. [§IV-D] 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.
  5. [Fig. 4] The caption of Fig. 4(b) does not identify which curve corresponds to which SNR scaling factor c. Please add a legend.
  6. [§IV-C] 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.

Circularity Check

0 steps flagged

No significant circularity: TFCDiff's claims are empirical benchmark results, externally anchored by held-out test sets and the SimEMG database.

full rationale

The paper's central claims are empirical benchmark measurements, not derivations that reduce to their own inputs. TFCDiff is trained on 33,294 synthesized QT/NST pairs and evaluated on a held-out 4,296-pair test set as well as the external SimEMG database never used in training (Section IV-B, Tables III and IX). No reported metric is a fitted parameter of the model, and the k-generation averaging is an evaluation protocol rather than a prediction. The DCT truncation to 50 Hz and the scaling bound eta from Eq. 9 are transparent preprocessing/hyperparameter choices, with truncation fidelity independently checked by SSD reconstruction error. The choice of c=150 is described as an empirical search ('We empirically find that the diffusion model yields the best performance when c is set to 150'), not as a derived result. Citations such as [30] for DCT-space diffusion and [33] for cross-domain feature fusion provide architectural inspiration but are not invoked as an unverified uniqueness theorem that forces the conclusion. The skeptic's band-limiting concern is a fairness/threat-to-validity issue for baseline comparison, not circularity, because the SimEMG generalization claim is externally grounded. The paper's own stated limitations (Section VI-D) acknowledge gaps such as poor performance under intensive EMG noise and lack of device validation, which further supports that no hidden circular step is being asserted as a derivation.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The method rests on standard DDPM machinery imported from the cited literature, plus domain assumptions about ECG spectral content and the realism of the synthetic noise model. The empirically tuned values (c=150, tau=1.75, lambda range, truncation at 1000 coefficients) are the main hand-chosen inputs; the paper discloses them and evaluates on held-out and external data, which keeps the circularity burden low. TFEM/TFE/TFF are architectural modules, not physically postulated entities.

free parameters (7)
  • SNR scaling factor c = 150
    Section III-D: 'We empirically find that the diffusion model yields the best performance when c is set to 150.' Sets the entire noise schedule; the ablation shows DCT-only diffusion is badly degraded without the rescaling/TFEM combination.
  • DC scaling percentile tau = 1.75 (yields eta ~= 3)
    Section III-C: chosen by 'empirical validation across multiple trials' to set the global normalization eta for all DCT inputs.
  • DCT truncation length = 1000 coefficients (~50 Hz cutoff)
    Section III-A: keeps the first [50/delta_f] coefficients on the assumption (after [38]) that ECG diagnostic content is below 50 Hz; justified by a truncation SSD check on 37590 segments.
  • noise intensity factor lambda = uniform in [0.2, 2]
    Section IV-B (Eq. 16): sets the SNR range of the fRMN synthesized train/test data and thus the difficulty distribution.
  • noise mixture weights r,m,n = r+m+n=1 (sampling distribution unspecified)
    Section IV-B: defines composite noise e = r*e1 + m*e2 + n*e3; how the weights are sampled is not stated in the text.
  • generation ensemble count k = 10 for the headline SOTA (tables also report k=1,3,5)
    Table III: the abstract's SOTA claim corresponds to k=10; k=1 is not SOTA on the synthesized benchmark. Multi-generation averaging is borrowed from DeScoD [18].
  • diffusion schedule endpoints = T=50, beta_1=1e-4, beta_T=0.5
    Section III-D: standard quadratic DDPM schedule; hand-chosen but conventional.
axioms (6)
  • standard math DDPM forward process, reverse process, and ELBO-derived L1 training objective (Eqs. 3-8)
    Imported from [23] and [39]; assumed valid without re-derivation.
  • domain assumption Truncated DCT coefficients of ECG are adequately modeled as diffusible data corrupted by unit Gaussian noise
    Core modeling choice; valid only with the SNR rescaling and TFEM according to the paper's own ablation (Table X).
  • domain assumption ECG diagnostic information lies below 50 Hz; coefficients beyond index 1000 are negligible
    Section III-A, after Zheng et al. [38]; supported by truncation SSD (1.66 +/- 3.30)e-3 on 37590 segments.
  • domain assumption QT recordings after bandpass/median filtering and piecewise-linear baseline removal within QRS-annotated segments are clean ground truth
    Section IV-B; every training target and test reference passes through this pipeline; fidelity is never validated against an independent reference.
  • domain assumption Amplitude-normalized linear mixtures of NST BW/MA/EM excerpts (Eq. 16) represent real ambulatory noise
    Section IV-B; defines the fRMN distribution for training and intra-dataset evaluation.
  • domain assumption Scaling bound eta computed from the DC component alone suffices to normalize the AC coefficients
    Section III-C and Fig. 3; AC coefficients span several orders of magnitude, so the normalization compresses small coefficients; the paper shows this is workable only in combination with the tuned schedule.

pith-pipeline@v1.3.0-alltime-deepseek · 20058 in / 25422 out tokens · 217814 ms · 2026-08-03T21:06:47.480597+00:00 · methodology

0 comments
read the original abstract

Ambulatory electrocardiogram (ECG) readings are prone to mixed noise from physical activities, including baseline wander (BW), muscle artifact (MA), and electrode motion artifact (EM). Developing a method to remove such complex noise and reconstruct high-fidelity signals is clinically valuable for diagnostic accuracy. However, denoising of multi-beat ECG segments remains understudied and poses technical challenges. To address this, we propose Time-Frequency Complementary Diffusion (TFCDiff), a novel approach that operates in the Discrete Cosine Transform (DCT) domain and uses the DCT coefficients of noisy signals as conditioning input. To refine waveform details, we incorporate Temporal Feature Enhancement Mechanism (TFEM) to reinforce temporal representations and preserve key physiological information. Comparative experiments on a synthesized dataset demonstrate that TFCDiff achieves state-of-the-art performance across five evaluation metrics. Furthermore, TFCDiff shows superior generalization on the unseen SimEMG Database, outperforming all benchmark models. Notably, TFCDiff processes raw 10-second sequences and maintains robustness under flexible random mixed noise (fRMN), enabling plug-and-play deployment in wearable ECG monitors for high-motion scenarios. Source code is available at https://github.com/Miroircivil/TFCDiff.

Figures

Figures reproduced from arXiv: 2511.16627 by Jie Min, Pengxin Li, Qingling Xia, Wang Li, Wei Liang, Yimin Zhou, Yirong Wang.

Figure 1
Figure 1. Figure 1: Schematic of the TFCDiff workflow. During training, the clean signal x0 is corrupted by the forward diffusion process q and the noise predictor, conditioned on the noisy observation x˜, learns to predict the added noise. During sampling, a random Gaussian noise xT is iteratively denoised via the reverse process p to reconstruct the denoised signal. which randomly selected one or more noise types and com￾bi… view at source ↗
Figure 2
Figure 2. Figure 2: Truncation of DCT coefficients for 10-s signals sampled at 360 Hz by retaining the first 1000 coefficients to preserve frequency content below 50 Hz. L(n) = N X−1 k=0 c(k)D(k) cos  (2n + 1)kπ 2N  , (2) where c(k) = (p 1/N, if k = 0. p 2/N, if k ̸= 0. Vital diagnostic information in normal ECG primarily lies in 0.5-50 Hz [38]. Since the full spectrum of D extends beyond this range, we truncate D to retain… view at source ↗
Figure 3
Figure 3. Figure 3: Distributions of the DC and AC components. Maroon lines indicate the 1st, 5th, 10th, 50th, 90th, 95th, and 99th percentiles for both components. estimation of the scaling bound η, using percentile-based truncation to mitigate the impact of outliers: η = max (|Pτ | , |P100−τ |), (9) where Pτ denotes the τ -th percentile of the DC component distribution [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Forward diffusion in the time domain and DCT domain using the same quadratic noise schedule. (a) Visualization of the forward diffusion process under varying SNR scaling factors. (b) SNR as a function of timesteps. making it difficult for the noise predictor to learn fine-grained denoising behavior. This often results in degraded detail recon￾struction in generated samples. Moreover, our global scaling of … view at source ↗
Figure 5
Figure 5. Figure 5: The architecture of the noise predictor, which comprises an U-Net backbone, TFE, TFF and other constitutive modules. while increasing computational cost. TFEM is composed of two modules: Temporal Feature Extraction (TFE) and Temporal Feature Fusion (TFF). In TFE, to convert a feature map from the DCT domain to the time domain, zero padding is first applied to restore its original length, followed by 1D IDC… view at source ↗
Figure 6
Figure 6. Figure 6: Dataset preprocessing. (a) An overview. (b) Three 10-second sequences sampled from the QT Database. (c) SNR distribution of the synthesized dataset. (d) SNR distribution of SimEMG Database. TABLE I DETAILS OF DATASETS Dataset QTDB NSTDB SimEMG Recordings 105 3 220 Length (second) 900 1800 30 Channel 2 2 1 Fs (Hz) 250 360 500 distinct sources, providing ground truth x0 for training and testing. It features … view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the denoised results on the synthesized dataset with representative samples from four different noise level intervals. TABLE II RECORDINGS SELECTED FOR TEST SET Database Recording ID MIT-BIH Arrhythmia Database sel123, sel233 MIT-BIH ST Change Database sel302, sel307 MIT-BIH Supraventricular Arrhythmia Database sel820, sel853 MIT-BIH Normal Sinus Rhythm Database sel16420, sel16795 European… view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the denoised results on the SimEMG Database. The top row displays two EMG-contaminated signals, and the bottom row presents the corresponding comparison between their ground truth and the denoised signals. TABLE VIII IMSNR (DB) COMPARISON RESULTS IN DIFFERENT NOISE LEVELS Models 0.2-0.6 0.6-1.0 1.0-1.5 1.5-2.0 FIR 5.29±3.15 6.66±3.41 6.69±3.41 6.77±3.44 IIR 4.50±3.51 6.33±3.52 6.55±3.42 6.… view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of the denoised results in the ablation study, with the yellow patches highlighting the ECG T-waves region. (a) TD. (b) TD+TFEM. (c) DCT. (d) DCT+TFEM. TABLE X EFFECTS OF DCT DIFFUSION AND TFEM TO THE BASELINE TIME-DOMAIN DIFFUSION MODEL. TD DCT TFEM SSD (au2 ) ↓ MAD (au) ↓ PRD (%) ↓ CosSim ↑ ImSNR (dB) ↑ ✓ 59.519±146.381 0.639±0.689 47.906±30.817 0.889±0.116 10.151±4.109 ✓ ✓ 49.960±166.622 0… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 4 linked inside Pith

  1. [1]

    Global Burden of Cardiovascular Diseases and Risks, 1990-2022,

    G. A. Mensah et al., “Global Burden of Cardiovascular Diseases and Risks, 1990-2022,”J. Am. Coll. Cardiol., vol. 82, no. 25, pp. 2350–2473, Dec. 2023

    1990

  2. [2]

    A survey on ECG analysis,

    S. Kaplan Berkaya, A. K. Uysal, E. Sora Gunal, S. Ergin, S. Gunal, and M. B. Gulmezoglu, “A survey on ECG analysis,”Biomed. Signal Process. Control, vol. 43, pp. 216–235, May 2018

    2018

  3. [3]

    A Systematic Review on the Use of Consumer-Based ECG Wearables on Cardiac Health Monitoring,

    R. Wang, S. C. M. Veera, O. Asan, and T. Liao, “A Systematic Review on the Use of Consumer-Based ECG Wearables on Cardiac Health Monitoring,”IEEE J. Biomed. Health Inform., vol. 28, no. 11, pp. 6525–6537, Nov. 2024

    2024

  4. [4]

    Advancing Sports Cardiology: Integrating Artificial Intelligence with Wearable Devices for Cardiovascular Health Manage- ment,

    X. Zheng et al., “Advancing Sports Cardiology: Integrating Artificial Intelligence with Wearable Devices for Cardiovascular Health Manage- ment,”ACS Appl. Mater . Interfaces, vol. 17, no. 12, pp. 17895–17920, Mar. 2025

    2025

  5. [5]

    State-of-the-art wearable sensors for cardiovascular health: a review,

    H. Xie, L. Yang, B. Jiang, Z. Huang, and Y . Lin, “State-of-the-art wearable sensors for cardiovascular health: a review,”Npj Cardiovasc. Health, vol. 2, no. 1, Oct. 2025, Art. no. 53

    2025

  6. [6]

    Opportunities and challenges of noise interference suppression algorithms for dynamic ECG signals in wearable devices: A review,

    J. Zhang et al., “Opportunities and challenges of noise interference suppression algorithms for dynamic ECG signals in wearable devices: A review,”Measurement, vol. 250, Jun. 2025, Art. no. 117067

    2025

  7. [7]

    Removal of noise from electrocardiogram using digital FIR and IIR filters with various meth- ods,

    K. S. Kumar, B. Yazdanpanah, and P. R. Kumar, “Removal of noise from electrocardiogram using digital FIR and IIR filters with various meth- ods,” inProc. IEEE Int. Conf. Commun. Signal Process., Melmaruvathur, India, Apr. 2015, pp. 0157–0162

    2015

  8. [8]

    Wavelet-domain group- sparse denoising method for ECG signals,

    C. Chen, M. Shu, S. Zhou, Z. Liu, and R. Liu, “Wavelet-domain group- sparse denoising method for ECG signals,”Biomed. Signal Process. Control, vol. 83, May 2023, Art. no. 104702

    2023

  9. [9]

    Power line noise and baseline wander removal from ECG signals using empirical mode decomposition and lifting wavelet transform technique,

    S. A. Malik, S. A. Parah, and B. A. Malik, “Power line noise and baseline wander removal from ECG signals using empirical mode decomposition and lifting wavelet transform technique,”Health Technol., vol. 12, no. 4, pp. 745–756, July 2022

    2022

  10. [10]

    Deep recurrent neural networks for ECG signal denoising,

    K. Antczak, “Deep recurrent neural networks for ECG signal denoising,” 2018, arXiv:1807.11551

    Pith/arXiv arXiv 2018

  11. [11]

    DeepFilter:An ECG baseline wander removal filter using deep learning techniques,

    F. P. Romero, D. C. Pi ˜nol, and C. R. V´azquez-Seisdedos, “DeepFilter:An ECG baseline wander removal filter using deep learning techniques,” Biomed. Signal Process. Control, vol. 70, 2021, Art. no. 102992

    2021

  12. [12]

    A lightweight U-Netmodel for denoising and noise localization of ECG signals,

    L. Hu, W. Cai, Z. Chen, and M. Wang, “A lightweight U-Netmodel for denoising and noise localization of ECG signals,”Biomed.Signal Process. Control, vol. 88, Feb. 2024, Art. no. 105504

    2024

  13. [13]

    Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders,

    H.-T. Chiang, Y .-Y . Hsieh, S.-W. Fu, K.-H. Hung, Y . Tsao, and S.- Y . Chien, “Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders,”IEEE Access, vol. 7, pp. 60806–60813, 2019

    2019

  14. [14]

    Convolutional block attention autoencoder for denoising electrocardiograms,

    W. Chorney, H. Wang, L. He, S. Lee, and L.-W. Fan, “Convolutional block attention autoencoder for denoising electrocardiograms,”Biomed. Signal Process. Control, vol. 86, Sep. 2023, Art. no.105242

    2023

  15. [15]

    Attention-Based Convolutional Denoising Au- toencoder for Two-Lead ECG Denoising and Arrhythmia Classification,

    P. Singh and A. Sharma, “Attention-Based Convolutional Denoising Au- toencoder for Two-Lead ECG Denoising and Arrhythmia Classification,” IEEE Trans. Instrum. Meas., vol. 71, pp. 1–10, 2022

    2022

  16. [16]

    Elimination of Random Mixed Noise in ECG Us- ing Convolutional Denoising Autoencoder With Transformer Encoder,

    M. Chen et al., “Elimination of Random Mixed Noise in ECG Us- ing Convolutional Denoising Autoencoder With Transformer Encoder,” IEEE J. Biomed. Health Inform., vol. 28, no. 4, pp. 1993–2004, Apr. 2024

    1993

  17. [17]

    An ECG Signal Denoising Method Using Conditional Generative Adversarial Net,

    X. Wang et al., “An ECG Signal Denoising Method Using Conditional Generative Adversarial Net,”IEEE J. Biomed. Health Inform., vol. 26, no. 7, pp. 2929–2940, July 2022

    2022

  18. [18]

    DeScoD-ECG: Deep Score- Based Diffusion Model for ECG Baseline Wander and Noise Removal,

    H. Li, G. Ditzler, J. Roveda, and A. Li, “DeScoD-ECG: Deep Score- Based Diffusion Model for ECG Baseline Wander and Noise Removal,” IEEE J. Biomed. Health Inform., vol. 28, no. 9, pp. 5081–5091, Sept. 2024

    2024

  19. [19]

    EDDM: A Novel ECG Denoising Method Using Dual-Path Diffusion Model,

    Z. Li et al., “EDDM: A Novel ECG Denoising Method Using Dual-Path Diffusion Model,”IEEE Trans. Instrum. Meas., vol. 74, 2025, Art no. 2509815

    2025

  20. [20]

    Leveraging conditional diffusion and pruning for optimal ECG signal denoising,

    B. A. Hassoon, S. Xiong, M. A. Hasson, and A. O. Abdulsalami, “Leveraging conditional diffusion and pruning for optimal ECG signal denoising,”Comput. Biol. Med., vol. 193, July 2025, Art. no. 110341

    2025

  21. [21]

    A Database of Simultaneously Recorded ECG Signals With and Without EMG Noise,

    V . Atanasoski et al., “A Database of Simultaneously Recorded ECG Signals With and Without EMG Noise,”IEEE Open J. Eng. Med. Biol., vol. 4, pp. 222–225, 2023

    2023

  22. [22]

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics,

    J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep Unsupervised Learning using Nonequilibrium Thermodynamics,” inProc. Int. Conf. Mach. Learn., Lille, France, Jul. 2015, pp. 2256–2265

    2015

  23. [23]

    Denoising diffusion probabilistic mod- els,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic mod- els,” inProc. Adv. Neural Inf. Process. Syst., Virtual, Jan. 2020, pp. 6840–6851

    2020

  24. [24]

    Generative Modeling by Estimating Gradients of the Data Distribution,

    Y . Song and S. Ermon, “Generative Modeling by Estimating Gradients of the Data Distribution,” inProc. Adv. Neural Inform. Process. Syst., Vancouver, Canada, Dec. 2019, pp. 11918-11930

    2019

  25. [25]

    Classifier-free diffusion guidance,

    J. Ho, and T. Salimans, “Classifier-free diffusion guidance,” 2022, arXiv:2207.12598. 14 IEEE TRANSACTIONS AND JOURNALS TEMPLATE

    Pith/arXiv arXiv 2022

  26. [26]

    DPM-solver: A fast ode solver for diffusion probabilistic modelsampling in around 10 steps

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-solver: A fast ode solver for diffusion probabilistic modelsampling in around 10 steps.” inProc. Adv. Neural Inform. Process. Syst., New Orleans, LA, USA, 2022, pp. 5775–5787

    2022

  27. [27]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inProc. Int. Conf. Learn. Represent., Kigali, Rwanda, May 2023, pp. 22698-22725

    2023

  28. [28]

    All are worth words: A ViT backbone for diffusion models,

    F. Bao et al., “All are worth words: A ViT backbone for diffusion models,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Vancouver, Canada, Jun. 2023, pp. 22669–22679

    2023

  29. [29]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProc. IEEE Int. Conf. Comput. Vis., Paris, France, Oct. 2023, pp. 4195–4205

    2023

  30. [30]

    DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space,

    M. Ning et al., “DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space,” inProc. Int. Conf. Mach. Learn., Van- couver, Canada, Jul. 2025, pp. 46498–46524

    2025

  31. [31]

    Fft-based dynamic token mixer for vision,

    Y . Tatsunami and M. Taki, “Fft-based dynamic token mixer for vision,” inProc. Conf. AAAI Artif. Intell., Vancouver, Canada, Feb. 2024, pp. 15328–15336

    2024

  32. [32]

    Efficient Frequency Domain- based Transformers for High-Quality Image Deblurring,

    L. Kong, J. Dong, J. Ge, M. Li, and J. Pan, “Efficient Frequency Domain- based Transformers for High-Quality Image Deblurring,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Vancouver, Canada, Jun. 2023, pp. 5886–5895

    2023

  33. [33]

    Complementary advantages: Exploiting cross-field frequency correlation for NIR-assisted image denoising,

    Y . Wang et al., “Complementary advantages: Exploiting cross-field frequency correlation for NIR-assisted image denoising,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Nashville, TN, USA, Jun. 2025, pp. 12679-12689

    2025

  34. [34]

    JPEG-LM: LLMs as Image Generators with Canonical Codec Representations,

    X. Han, M. Ghazvininejad, P. W. Koh, and Y . Tsvetkov, “JPEG-LM: LLMs as Image Generators with Canonical Codec Representations,” 2024, arXiv:2408.08459

    Pith/arXiv arXiv 2024

  35. [35]

    Fourier Image Transformer,

    T.-O. Buchholz and F. Jug, “Fourier Image Transformer,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., New Orleans, LA, USA, Jun. 2022, pp. 1846–1854

    2022

  36. [36]

    Discrete Cosine Transform,

    N. Ahmed, T. Natarajan, and K. R. Rao, “Discrete Cosine Transform,” IEEE Trans. Comput., vol. C–23, no. 1, pp. 90–93, Jan. 1974

    1974

  37. [37]

    The JPEG still picture compression standard,

    G. K. Wallace, “The JPEG still picture compression standard,”IEEE Trans. Consum. Electron., vol. 38, no. 1, pp. 18-34, Feb. 1992

    1992

  38. [38]

    A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients,

    J. Zheng, J. Zhang, S. Danioko, H. Yao, H. Guo, and C. Rakovski, “A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients,”Sci. Data, vol. 7, Jan. 2020, Art. no. 48

    2020

  39. [39]

    Understanding Diffusion Models: A Unified Perspective,

    C. Luo, “Understanding Diffusion Models: A Unified Perspective,” 2022, arXiv:2208.11970

    Pith/arXiv arXiv 2022

  40. [40]

    Wavegrad: Estimating gradients for waveform generation,

    N. Chen, Y . Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “Wavegrad: Estimating gradients for waveform generation,” inProc. Int. Conf. Learn. Represent., Vienna, Austria, May 2021, pp. 13188-13202

    2021

  41. [41]

    U-Net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” inProc. Int. Conf. Med. Image Comput. Comput.-Assist. Interv., Munich, Germany, Oct. 2015, pp. 234–241

    2015

  42. [42]

    Saharia, J

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, ”Image Super-Resolution via Iterative Refinement,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 4, pp. 4713-4726, Apr. 2023

    2023

  43. [43]

    FiLM: visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville, “FiLM: visual reasoning with a general conditioning layer,” inProc. Conf. AAAI Artif. Intell., New Orleans, LA, USA, Feb. 2018, pp. 3942-3951

    2018

  44. [44]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” inAdv. Neural inform. Process. Syst., Long Beach, CA, USA, Dec. 2017, pp. 6000-6010

    2017

  45. [45]

    A database for evaluation of algorithms for measurement of QT and other wave-form intervals in the ECG,

    P. Laguna, R. G. Mark, A. Goldberg, and G. B. Moody, “A database for evaluation of algorithms for measurement of QT and other wave-form intervals in the ECG,” inProc. IEEE Comput. Cardiol., Lund, Sweden, 1997, pp. 673–676

    1997

  46. [46]

    A noise stress test for arrhythmia detectors,

    G. Moody, W. Muldrow, and R. Mark, “A noise stress test for arrhythmia detectors,” inProc. Comput. Cardiol., 1984, pp. 381–384

    1984

  47. [47]

    PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals,

    A. L. Goldberger et al., “PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals,”Circulation, vol. 101, no. 23, pp. e215–e220, 2000

    2000

  48. [48]

    Y . Tao, B. Xu, and Y . Zhang, ”Refined Self-Attention Transformer Model for ECG-Based Arrhythmia Detection,”IEEE Trans. Instrum. Meas., vol. 73, pp. 1-14, 2024, Art no. 4007314

    2024

  49. [49]

    Attention-refined U-Net with Skip Connections for Effective Brain Tumor Segmentation from MRI Images,

    A. F. M. M. Rahman and Md. A. Hossain, “Attention-refined U-Net with Skip Connections for Effective Brain Tumor Segmentation from MRI Images,” inInt. Conf. Comput. Inform. Technol., Cox’s Bazar, Bangladesh, Dec. 2023, pp. 1-6

    2023

  50. [50]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu, “Squeeze-and-excitation networks,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Salt Lake City, UT, USA, Jun. 2018, pp. 7132–7141

    2018

  51. [51]

    Dynamic convolution:Attention over convolution kernels,

    Y . Chen, X. Dai, M. Liu, D. Chen, L. Yuan, and Z. Liu, “Dynamic convolution:Attention over convolution kernels,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Seattle, W A, USA, Jun. 2020, pp. 11030–11039

    2020

  52. [52]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” inProc. Int. Conf. Mach. Learn., Sydney, Australia, Aug. 2017, pp. 3319–3328

    2017