TF-TransUNet1D: Time-Frequency Guided Transformer U-Net for Robust ECG Denoising in Digital Twin

Lei Li; Shijie Wang

REVIEW 4 major objections 6 minor 18 references

TF-TransUNet1D: Time-Frequency Guided Transformer U-Net for Robust ECG Denoising in Digital Twin

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read 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.

desk verdict Plausible engineering combo of Transformer U-Net and spectral loss; claims of consistent superiority outrun the single-condition, no-error-bar evidence. read the letter →

arxiv 2508.20398 v1 pith:PPFOJKJY submitted 2025-08-28 cs.LG cs.AI

Shijie Wang , Lei Li This is my paper · ORCID

classification cs.LGcs.AI

keywords ECGdenoisingTransUNettime-frequencylossspectralfidelitycardiacdigitaltwinself-attentionsignal-to-noiseratioimprovement

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the 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.

What 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.

The 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.

Minor: 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.

Who 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.

My 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.

Referee Report

4 major / 6 minor

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.

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 (4)

[§3.2, Table 1] 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.
[§3.1 Implementation and §2.3 (Eq. 5)] 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.
[§1 and Table 1] 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.
[§3.1 Noisy Signal Synthesis and Preprocessing] 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.

minor comments (6)

[§2.2] Equation (1) is referenced but not displayed; please number and show all equations, including Eqs. (2) and (3) if used.
[§3.2 and Fig. 2] 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.
[§3.2] '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).
[§3.1] The PRD metric is defined but never reported. Either report PRD in the results or remove it from the evaluation metrics section.
[§4] 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.
[§3.1] Minor typo: 'NVIDIA GeForce RTX 4090GPU' should have a space before 'GPU'. Please proofread the manuscript for similar spacing and capitalization issues.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported metrics are held-out test measurements, not derived from the loss by construction.

full rationale

TF-TransUNet1D's derivation chain is not circular. The model is trained to minimize L_total = w_time * SmoothL1(y, ŷ) + w_spectral * FFT-magnitude MSE (Eqs. 4–5) and then evaluated on held-out, subject-disjoint test segments using MAE, PCC, and SNRI. The reported numbers (MAE 0.1285, PCC 0.9540, SNRI 13.36 at 0 dB mixed noise) are measured test-set results, not quantities obtained by plugging fitted constants back into the same equations. The validation set is used only to tune hyperparameters (loss weights, learning rate, early stopping), a standard procedure that does not force the test metrics. The only self-citations (refs. 1 and 2 by co-author Lei Li) appear in the introduction as motivation for cardiac digital twins; they are not used as an external uniqueness or correctness argument for the denoising architecture. No equation defines the claimed output in terms of the input, no fitted parameter is renamed as a prediction, and no 'uniqueness' theorem is imported. Concerns about baseline breadth, error bars, and synthetic-noise realism are statistical validity/correctness issues, not circularity. Under the stated criteria, the finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its free parameters are loss weights and architecture hyperparameters that are tuned but not reported, which weakens reproducibility. The key domain assumption is that synthetic NSTDB noise is a faithful proxy for real-world ECG noise.

free parameters (4)

loss weight w_time = not reported
Introduced in Eq. (5); value tuned on the validation set, not reported in the paper.
loss weight w_spectral = not reported
Introduced in Eq. (5); value tuned on the validation set, not reported.
Smooth L1 threshold beta = not reported
Chosen for convergence stability and robustness, but no value given in Section 2.3.
Transformer depth N, heads H, embedding dimension d = not reported
Standard model hyperparameters that significantly affect capacity and results, but not specified in Section 2.2.

assumptions (4)

standard math FFT magnitude comparison is a valid measure of spectral fidelity for ECG signals.
Used in Eq. (4) to define L_spectral; relies on standard Fourier analysis.
domain assumption MIT-BIH clean recordings mixed with NSTDB noise at fixed SNRs represent realistic ECG corruption for digital twin pipelines.
Introduced in Section 3.1 'Noisy Signal Synthesis and Preprocessing'; the entire evaluation depends on this proxy.
domain assumption Test segments from subjects not seen during training ensure fair generalization assessment.
Stated in Section 3.1; the paper does not provide patient-level split details or verify no leakage.
ad hoc to paper A weighted sum of time and spectral losses with manually chosen weights is an effective training objective.
Eq. (5) defines the total loss; the weights are selected by validation tuning and are central to the method.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TF-TransUNet1D: Time-Frequency Guided Transformer U-Net for Robust ECG Denoising in Digital Twin." pith.science (2026). https://pith.science/paper/PPFOJKJY

@misc{pith2026250820398,
  author       = {Pith},
  title        = {Pith review of: TF-TransUNet1D: Time-Frequency Guided Transformer U-Net for Robust ECG Denoising in Digital Twin},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPFOJKJY}},
  note         = {Machine review of arXiv:2508.20398}
}

read the original abstract

Electrocardiogram (ECG) signals serve as a foundational data source for cardiac digital twins, yet their diagnostic utility is frequently compromised by noise and artifacts. To address this issue, we propose TF-TransUNet1D, a novel one-dimensional deep neural network that integrates a U-Net-based encoder-decoder architecture with a Transformer encoder, guided by a hybrid time-frequency domain loss. The model is designed to simultaneously capture local morphological features and long-range temporal dependencies, which are critical for preserving the diagnostic integrity of ECG signals. To enhance denoising robustness, we introduce a dual-domain loss function that jointly optimizes waveform reconstruction in the time domain and spectral fidelity in the frequency domain. In particular, the frequency-domain component effectively suppresses high-frequency noise while maintaining the spectral structure of the signal, enabling recovery of subtle but clinically significant waveform components. We evaluate TF-TransUNet1D using synthetically corrupted signals from the MIT-BIH Arrhythmia Database and the Noise Stress Test Database (NSTDB). Comparative experiments against state-of-the-art baselines demonstrate consistent superiority of our model in terms of SNR improvement and error metrics, achieving a mean absolute error of 0.1285 and Pearson correlation coefficient of 0.9540. By delivering high-precision denoising, this work bridges a critical gap in pre-processing pipelines for cardiac digital twins, enabling more reliable real-time monitoring and personalized modeling.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

[1]

Medical Image Analysis 101, 103472 (2025)

Li, L., Smith, H., Lyu, Y., Camps, J., Qian, S., Rodriguez, B., Banerjee, A., Grau, V.: Per- sonalized topology -informed localization of standard 12 -lead ECG electrode placement from incomplete cardiac MRIs for efficient cardiac digital twins. Medical Image Analysis 101, 103472 (2025)

work page 2025
[2]

IEEE Transactions on Medical Imaging 43(7), 2466–2478 (2024)

Li, L., Camps, J., Wang, Z.J., Beetz, M., Banerjee, A., Rodriguez, B., Grau, V.: Toward enabling cardiac digital twins of myocardial infarction using deep computational models for inverse inference. IEEE Transactions on Medical Imaging 43(7), 2466–2478 (2024)

work page 2024
[3]

ISA Transactions 114, 251–262 (2021)

Kumar, A., Sharma, A., Pachori, R.B.: Stationary wavelet transform based ECG signal de- noising method. ISA Transactions 114, 251–262 (2021)

work page 2021
[4]

In: Proceedings of the 2012 Computing in Cardiology Conference (2012)

Romero, I., Geng, D., Berset, T.: Adaptive filtering in ECG denoising: A comparative study. In: Proceedings of the 2012 Computing in Cardiology Conference (2012)

work page 2012
[5]

Biomedical Signal Processing and Control 7(5), 481 –489 (2012)

Kabir, M.A., Shahnaz, C.: Denoising of ECG signals based on noise reduction algorithms in EMD and wavelet domains. Biomedical Signal Processing and Control 7(5), 481 –489 (2012)

work page 2012
[6]

Biomedical Signal Processing and Control 69, 102903 (2021)

Sraitih, M., Jabrane, Y.: A denoising performance comparison based on ECG signal decom- position and local means filtering. Biomedical Signal Processing and Control 69, 102903 (2021)

work page 2021
[7]

In: Proceedings of the 2006 International Conference of the IEEE Engineer- ing in Medicine and Biology Society

Weng, B., Blanco-Velasco, M., Barner, K.E.: ECG denoising based on the empirical mode decomposition. In: Proceedings of the 2006 International Conference of the IEEE Engineer- ing in Medicine and Biology Society

work page 2006
[8]

IEEE Sensors Journal 22(23), 22725–22733 (2022)

Li, C., Zhang, X., Wang, Y., Liu, H., Chen, Z., Liu, J.: ECG denoising method based on an improved VMD algorithm. IEEE Sensors Journal 22(23), 22725–22733 (2022)

work page 2022

Show all 18 references

[9]

-Y.: Noise reduc- tion in ECG signals using fully convolutional denoising autoencoders

Chiang, H.-T., Hsieh, Y.-Y., Fu, S.-W., Hung, K.-H., Tsao, Y., Chien, S. -Y.: Noise reduc- tion in ECG signals using fully convolutional denoising autoencoders. IEEE Access 7, 60806–60813 (2019)

2019
[10]

arXiv preprint arXiv:1807.11551 (2018)

Antczak, K.: Deep recurrent neural networks for ECG signal denoising. arXiv preprint arXiv:1807.11551 (2018)

2018 arXiv
[11]

Biomedical Signal Processing and Control 63, 102225 (2021)

Dasan, E., Panneerselvam, I.: A novel dimensionality reduction approach for ECG signal via convolutional denoising autoencoder with LSTM. Biomedical Signal Processing and Control 63, 102225 (2021)

2021
[12]

IEEE Journal of Bio- medical and Health Informatics (2023)

Li, H., Zhou, Y., Liu, H., Peng, Y., Zhang, J., Li, Y., Wang, Y.: Descod -ECG: Deep score- based diffusion model for ECG baseline wander and noise removal. IEEE Journal of Bio- medical and Health Informatics (2023)

2023
[13]

IEEE Transactions on Instrumentation and Measurement 71, 1–10 (2022)

Singh, P., Sharma, A.: Attention -based convolutional denoising autoencoder for two -lead ECG denoising and arrhythmia classification. IEEE Transactions on Instrumentation and Measurement 71, 1–10 (2022). Art. no. 4007710

2022
[14]

IEEE Transactions on Instrumentation and Measurement 72, 1–11 (2023)

Hou, Y., Liu, R., Shu, M., Xie, X., Chen, C.: Deep neural network denoising model based on sparse representation algorithm for ECG signal. IEEE Transactions on Instrumentation and Measurement 72, 1–11 (2023). Art. no. 2507711

2023
[15]

arXiv preprint arXiv:2407.11065 (2024)

Zhu, D., Chhabra, V.K., Khalili, M.M.: ECG Signal Denoising Using Multi-scale Patch Em- bedding and Transformers. arXiv preprint arXiv:2407.11065 (2024)

2024 arXiv
[16]

Electronics 12(7), 1606 (2023)

Lin, H., Liu, R., Liu, Z.: ECG signal denoising method based on disentangled autoencoder. Electronics 12(7), 1606 (2023)

2023
[17]

Phys- iological Measurement 42(11), 115002 (2021)

Qiu, L., et al.: Two-stage ECG signal denoising based on deep convolutional network. Phys- iological Measurement 42(11), 115002 (2021)

2021
[18]

Biomedical Signal Processing and Control 50, 62–71 (2019)

Mourad, N.: ECG denoising algorithm based on group sparsity and singular spectrum anal- ysis. Biomedical Signal Processing and Control 50, 62–71 (2019)

2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Medical Image Analysis 101, 103472 (2025)

Li, L., Smith, H., Lyu, Y., Camps, J., Qian, S., Rodriguez, B., Banerjee, A., Grau, V.: Per- sonalized topology -informed localization of standard 12 -lead ECG electrode placement from incomplete cardiac MRIs for efficient cardiac digital twins. Medical Image Analysis 101, 103472 (2025)

work page 2025

[2] [2]

IEEE Transactions on Medical Imaging 43(7), 2466–2478 (2024)

Li, L., Camps, J., Wang, Z.J., Beetz, M., Banerjee, A., Rodriguez, B., Grau, V.: Toward enabling cardiac digital twins of myocardial infarction using deep computational models for inverse inference. IEEE Transactions on Medical Imaging 43(7), 2466–2478 (2024)

work page 2024

[3] [3]

ISA Transactions 114, 251–262 (2021)

Kumar, A., Sharma, A., Pachori, R.B.: Stationary wavelet transform based ECG signal de- noising method. ISA Transactions 114, 251–262 (2021)

work page 2021

[4] [4]

In: Proceedings of the 2012 Computing in Cardiology Conference (2012)

Romero, I., Geng, D., Berset, T.: Adaptive filtering in ECG denoising: A comparative study. In: Proceedings of the 2012 Computing in Cardiology Conference (2012)

work page 2012

[5] [5]

Biomedical Signal Processing and Control 7(5), 481 –489 (2012)

Kabir, M.A., Shahnaz, C.: Denoising of ECG signals based on noise reduction algorithms in EMD and wavelet domains. Biomedical Signal Processing and Control 7(5), 481 –489 (2012)

work page 2012

[6] [6]

Biomedical Signal Processing and Control 69, 102903 (2021)

Sraitih, M., Jabrane, Y.: A denoising performance comparison based on ECG signal decom- position and local means filtering. Biomedical Signal Processing and Control 69, 102903 (2021)

work page 2021

[7] [7]

In: Proceedings of the 2006 International Conference of the IEEE Engineer- ing in Medicine and Biology Society

Weng, B., Blanco-Velasco, M., Barner, K.E.: ECG denoising based on the empirical mode decomposition. In: Proceedings of the 2006 International Conference of the IEEE Engineer- ing in Medicine and Biology Society

work page 2006

[8] [8]

IEEE Sensors Journal 22(23), 22725–22733 (2022)

Li, C., Zhang, X., Wang, Y., Liu, H., Chen, Z., Liu, J.: ECG denoising method based on an improved VMD algorithm. IEEE Sensors Journal 22(23), 22725–22733 (2022)

work page 2022