pith. sign in

arxiv: 2605.03183 · v2 · pith:YU7OVAEMnew · submitted 2026-05-04 · 💻 cs.LG · eess.SP

Enhancing AI-Based ECG Delineation with Deep Learning Denoising Techniques

Jeff Breeding-Allison , Emil Walleser This is my paper

Pith reviewed 2026-05-20 23:28 UTC · model grok-4.3

classification 💻 cs.LG eess.SP
keywords canine ECGECG denoisingautoencoderdeep learningsignal preprocessingnoise reductionAI delineation
0
0 comments X

The pith

An autoencoder neural network effectively denoises canine ECG signals while preserving features for AI-based delineation.

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

The paper proposes training an autoencoder neural network to reconstruct clean cardiac signals from noisy canine ECG inputs as a preprocessing step. Noise from respiration, muscle activity, poor contact, and external artifacts often obscures key waveform details, and classical filters or wavelets struggle to handle all patterns without distorting important morphology. The model aims to suppress interference across both noisy and clean recordings while keeping diagnostically relevant features intact. This matters for making automated AI delineation more reliable in veterinary ECG analysis.

Core claim

The central claim is that an autoencoder-based neural network trained to reconstruct clean cardiac signals from noisy inputs enables effective noise reduction for canine ECGs without degrading diagnostically important waveforms. The approach shows strong performance across noisy and clean recordings, demonstrating robustness to varying signal conditions and suitability for downstream delineation tasks.

What carries the argument

Autoencoder neural network trained to reconstruct clean signals from noisy ECG inputs, serving as preprocessing to suppress interference while preserving morphological features.

If this is right

  • Denoised signals become suitable inputs for AI models performing ECG delineation.
  • The method maintains performance regardless of whether recordings contain high or low noise levels.
  • It provides a practical alternative when classical filtering fails to handle mixed interference sources.
  • Robustness across conditions supports use in varied real-world veterinary recording setups.

Where Pith is reading between the lines

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

  • Combining the denoising step with full end-to-end AI pipelines could raise overall accuracy in automated canine heart analysis.
  • The same training approach might transfer to denoising other physiological signals with comparable artifact problems.
  • Measuring end-to-end effects on specific outputs like P-wave or T-wave detection would test the preprocessing value more precisely.

Load-bearing premise

The autoencoder successfully suppresses diverse noise patterns while preserving morphological features critical for accurate ECG delineation without introducing new artifacts that degrade downstream AI performance.

What would settle it

Comparing the accuracy of an AI-based delineator on the same canine ECG recordings processed with and without the autoencoder denoising step to check for measurable gains or introduced distortions.

Figures

Figures reproduced from arXiv: 2605.03183 by Emil Walleser, Jeff Breeding-Allison.

Figure 1
Figure 1. Figure 1: An ECG lead with baseline wander. + = view at source ↗
Figure 1
Figure 1. Figure 1: An ECG lead with baseline wander. data. To mimic baseline wander, we synthetically add a sinusoidal component to an ECG segment, as shown in [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Baseline wander noising process. • the frequency, and • the phase shift. For each added sinusoidal component, these parameters are randomly sampled from a predefined search space. To emulate baseline wander observed in real-world ECG record￾ings, the amplitude is drawn from a narrow range informed by empirical ECG data, while the frequency is selected from a range consistent with expected respiratory rates… view at source ↗
Figure 2
Figure 2. Figure 2: Baseline wander noising process. irregular fluctuations superimposed on the ECG signal. Although measures such as improv￾ing patient comfort and relaxation can reduce the prevalence of this type of noise, complete elimination is rarely achievable in practice. Muscle-related interference can obscure clini￾cally relevant features, including low-amplitude waves such as the P-wave or T-wave, and in some cases … view at source ↗
Figure 3
Figure 3. Figure 3: A delineated base ECG. 4. Results In the denoising-delineation ECG pipeline noisy ECG → denoised ECG → delineated ECG, there are two natural points at which denoising model performance can be evaluated: im￾mediately after denoising and after downstream ECG delineation. At the first evaluation point, we quantify changes in signal quality using signal-to-noise ratio (SNR) and the ad￾ditional noise metrics de… view at source ↗
Figure 3
Figure 3. Figure 3: A delineated base ECG. point, we quantify changes in signal quality using signal-to-noise ratio (SNR) and the ad￾ditional noise metrics described above. As baselines, we compare the performance of our deep learning models against classical signal filtering techniques. At the second evaluation point, we assess performance using the delineation metrics introduced previously. Because the primary goal of denoi… view at source ↗
Figure 4
Figure 4. Figure 4: The delineated noised base ECG view at source ↗
Figure 4
Figure 4. Figure 4: The delineated noised base ECG [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The elgendi2010 filtered noised ECG view at source ↗
Figure 5
Figure 5. Figure 5: The elgendi2010 filtered noised ECG [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The delineated elgendi2010 filtered noised ECG. In contrast, applying our autoencoder-based denoising model to the noised ECG produces the signal shown in view at source ↗
Figure 6
Figure 6. Figure 6: The delineated elgendi2010 filtered noised ECG. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The autoencoder cleaned noised ECG view at source ↗
Figure 7
Figure 7. Figure 7: The autoencoder cleaned noised ECG. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The delineated autoencoder cleaned noised ECG. denoising approach over the selected classical filter for improving downstream ECG delin￾eation. 4.1. Denoising performance metrics. Across both noisy and clean ECG signals, our autoencoder-based denoising model consistently produces ECG inputs that remain suitable for downstream delineation. While certain classical filtering techniques may slightly out￾perfor… view at source ↗
Figure 8
Figure 8. Figure 8: The delineated autoencoder cleaned noised ECG. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
read the original abstract

Evaluating canine electrocardiograms (ECGs) is challenging due to noise that can obscure clinically relevant cardiac electrical activity. Common sources of interference include respiration, muscle activity, poor lead contact, and external electrical artifacts. Classical signal denoising techniques, such as filtering and wavelet-based methods, struggle to suppress diverse noise patterns while preserving morphological features critical for accurate ECG delineation. We propose an autoencoder-based neural network model and training strategy for ECG denoising as a preprocessing step for canine ECG analysis. The model is trained to reconstruct clean cardiac signals from noisy inputs, enabling effective noise reduction without degrading diagnostically important waveforms. Our approach demonstrates strong performance across both noisy and clean ECG recordings, indicating robustness to varying signal conditions and suitability for downstream delineation tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes an autoencoder-based neural network trained to reconstruct clean canine ECG signals from noisy inputs as a preprocessing step for AI-based ECG delineation. It claims the approach suppresses diverse noise sources (respiration, muscle activity, artifacts) while preserving morphological features, yielding strong performance on both noisy and clean recordings and indicating suitability for downstream delineation tasks.

Significance. If the central empirical claims were supported by quantitative validation, the work would provide a data-driven alternative to classical filtering and wavelet methods for noisy veterinary ECGs, potentially improving robustness of AI delineators in real-world conditions where noise is prevalent.

major comments (2)
  1. [Abstract] Abstract: The claim that the model 'demonstrates strong performance across both noisy and clean ECG recordings, indicating robustness ... and suitability for downstream delineation tasks' supplies no quantitative metrics (SNR, MSE, onset/offset MAE, F1 scores), datasets, baselines, or validation details. This prevents any assessment of whether the denoising step actually improves or preserves delineation accuracy.
  2. [Abstract / Methods] The argument for suitability rests on the unverified assumption that the autoencoder suppresses noise while exactly preserving P/QRS/T wave shapes and boundaries without introducing artifacts that could increase boundary errors. No end-to-end experiments measuring delineation metrics before versus after denoising on canine data are described.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We agree that strengthening the quantitative support in the abstract and adding explicit end-to-end validation would improve the manuscript. Below we respond point by point and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that the model 'demonstrates strong performance across both noisy and clean ECG recordings, indicating robustness ... and suitability for downstream delineation tasks' supplies no quantitative metrics (SNR, MSE, onset/offset MAE, F1 scores), datasets, baselines, or validation details. This prevents any assessment of whether the denoising step actually improves or preserves delineation accuracy.

    Authors: We accept that the original abstract was too concise and omitted concrete numbers. In the revised manuscript we have expanded the abstract to report the key quantitative results obtained on the held-out canine test set: mean SNR improvement, MSE, and the downstream delineation F1 scores together with onset/offset MAE for P/QRS/T waves. The datasets (canine ECGs with synthetic and real noise) and the classical baselines used for comparison are now explicitly named in the abstract and detailed in the Methods section. revision: yes

  2. Referee: [Abstract / Methods] The argument for suitability rests on the unverified assumption that the autoencoder suppresses noise while exactly preserving P/QRS/T wave shapes and boundaries without introducing artifacts that could increase boundary errors. No end-to-end experiments measuring delineation metrics before versus after denoising on canine data are described.

    Authors: We agree that an explicit before-versus-after comparison on the downstream delineation task is necessary to substantiate the claim of suitability. We have added a new subsection in the Results that applies a fixed AI delineator to the original noisy recordings and to the autoencoder-denoised versions of the same canine recordings. We report the resulting changes in onset/offset MAE and F1 scores for each wave, confirming that boundary errors do not increase and in most cases decrease after denoising. These experiments directly address the concern about morphological preservation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on standard training and evaluation

full rationale

The paper presents an autoencoder trained to reconstruct clean ECG signals from noisy inputs as a preprocessing step, with performance claims framed as empirical outcomes on noisy and clean recordings. No equations, fitted parameters renamed as predictions, or self-referential definitions appear in the provided abstract or description. The suitability for downstream delineation is asserted based on observed robustness rather than any derivation that reduces to its own inputs by construction. This is a standard ML pipeline without load-bearing self-citations, uniqueness theorems, or ansatzes smuggled via prior work. Any gap in explicit end-to-end validation is a matter of experimental completeness, not circularity in the logical chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, preventing extraction of specific free parameters, axioms, or invented entities. The approach implicitly relies on standard deep-learning assumptions about signal reconstruction and feature preservation.

pith-pipeline@v0.9.0 · 5646 in / 986 out tokens · 54219 ms · 2026-05-20T23:28:04.146660+00:00 · methodology

Review history (3 revisions) →

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

16 extracted references · 16 canonical work pages

  1. [1]

    Wu, and Richard T

    Vennela Avula, Katherine C. Wu, and Richard T. Carrick. Clinical Applications, Methodology, and Scientific Reporting of Electrocardiogram Deep-Learning Models: A Systematic Review.JACC: Advances, 2(10):100686, 2023

    work page 2023

  2. [2]

    Raghavendra Badiger and Prabhakar M. Ascnet-ecg: Deep autoencoder based atten- tion aware skip connection network for ecg filtering.International Journal of Engi- neering Trends and Technology, 71(2):382–398, February 2023

    work page 2023

  3. [3]

    Cardoso, Lisa Bedin, Josselin Duchateau, R´ emi Dubois, and Eric Moulines

    Gabriel V. Cardoso, Lisa Bedin, Josselin Duchateau, R´ emi Dubois, and Eric Moulines. Bayesian ecg reconstruction using denoising diffusion generative models, 2023

    work page 2023

  4. [4]

    Noise reduction in ecg signals using fully convolutional denoising autoencoders

    Hsin-Tien Chiang, Yi-Yen Hsieh, Szu-Wei Fu, Kuo-Hsuan Hung, Yu Tsao, and Shao-Yi Chien. Noise reduction in ecg signals using fully convolutional denoising autoencoders. IEEE Access, 7:60806–60813, 01 2019

    work page 2019

  5. [5]

    Pulsenet: Deep learning ecg-signal classifi- cation using random augmentation policy and continous wavelet transform for canines, 2023

    Andre Dourson, Roberto Santilli, Federica Marchesotti, Jennifer Schneiderman, Oliver Roman Stiel, Fernando Junior, Michael Fitzke, Norbert Sithirangathan, Emil Walleser, Xiaoli Qiao, and Mark Parkinson. Pulsenet: Deep learning ecg-signal classifi- cation using random augmentation policy and continous wavelet transform for canines, 2023

    work page 2023

  6. [6]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

    work page 2020

  7. [7]

    Ecgassess: A python-based tool- box to assess ecg lead signal quality.Frontiers in Digital Health, 4, 05 2022

    Linus Kramer, Carlo Menon, and Mohamed Elgendi. Ecgassess: A python-based tool- box to assess ecg lead signal quality.Frontiers in Digital Health, 4, 05 2022

    work page 2022

  8. [8]

    Descod-ecg: Deep score-based diffusion model for ecg baseline wander and noise removal.IEEE Journal of Biomedical and Health Informatics, 28(9):5081–5091, September 2024

    Huayu Li, Gregory Ditzler, Janet Roveda, and Ao Li. Descod-ecg: Deep score-based diffusion model for ecg baseline wander and noise removal.IEEE Journal of Biomedical and Health Informatics, 28(9):5081–5091, September 2024. 17 JEFF BREEDING-ALLISON AND EMIL W ALLESER

    work page 2024

  9. [9]

    Lau, Jan C

    Dominique Makowski, Tam Pham, Zen J. Lau, Jan C. Brammer, Fran¸ cois Lespinasse, Hung Pham, Christopher Sch¨ olzel, and S. H. Annabel Chen. NeuroKit2: A python tool- box for neurophysiological signal processing.Behavior Research Methods, 53(4):1689– 1696, feb 2021

    work page 2021

  10. [10]

    Adewole, Hammed A

    Nehemiah Musa, Abdulsalam Ya’u Gital, Nahla Aljojo, Haruna Chiroma, Kayode S. Adewole, Hammed A. Mojeed, Nasir Faruk, Abubakar Abdulkarim, Ifada Emmanuel, Yusuf Y. Folawiyo, James A. Ogunmodede, Abdukareem A. Oloyede, Lukman A. Ola- woyin, Ismaeel A. Sikiru, and Ibrahim Katb. A systematic review and meta-data analysis on the applications of deep learning ...

    work page 2023

  11. [11]

    Pytorch: An imperative style, high-performance deep learning library, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Des- maison, Andreas K¨ opf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-perf...

    work page 2019

  12. [12]

    Deepfilter: an ecg baseline wander removal filter using deep learning techniques, 2021

    Francisco Perdigon Romero, David Castro Pi˜ nol, and Carlos Rom´ an V´ azquez Seisde- dos. Deepfilter: an ecg baseline wander removal filter using deep learning techniques, 2021

    work page 2021

  13. [13]

    II Edition

    Roberto Santilli, Sidney Mo¨ ıse, Romain Pariaut, and Manuela Perego.Electrocardio- graphy of the Dog and Cat: Diagnosis of Arrhythmias. II Edition. Edra, 2 edition, 2018

    work page 2018

  14. [14]

    Sabarimalai Manikandan

    Udit Satija, Barathram Ramkumar, and M. Sabarimalai Manikandan. A review of signal processing techniques for electrocardiogram signal quality assessment.IEEE Reviews in Biomedical Engineering, 11:36–52, 2018

    work page 2018

  15. [15]

    Denoising autoencoder for eletrocardiogram signal enhancement.Journal of Medical Imaging and Health Infor- matics, 5:1804–1810, 12 2015

    Peng Xiong, Hongrui Wang, Ming Liu, and Xiuling Liu. Denoising autoencoder for eletrocardiogram signal enhancement.Journal of Medical Imaging and Health Infor- matics, 5:1804–1810, 12 2015

    work page 2015

  16. [16]

    An ecg denoising method based on the generative adversarial residual network.Computational and Mathematical Methods in Medicine, 2021:1–23, 04 2021

    Bingxin Xu, Ruixia Liu, Minglei Shu, Xiaoyi Shang, and Yinglong Wang. An ecg denoising method based on the generative adversarial residual network.Computational and Mathematical Methods in Medicine, 2021:1–23, 04 2021. 18 ENHANCING AI-BASED ECG DELINEATION WITH DEEP LEARNING DENOISING TECHNIQUES Appendices A Denoising performance tables Denoiser Mean Std ...

    work page 2021