Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Robust and Generalizable Atrial Fibrillation Detection from ECG Using Time-Frequency Fusion and Supervised Contrastive Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A dual-branch ECG model that fuses time and frequency features with supervised contrastive learning outperforms six published AF detectors on both within-dataset and cross-dataset tests.

desk verdict The architecture is sensible and the intra-dataset results are solid, but the headline cross-dataset gains are inflated by selecting the contrastive weight on the target test set and by filtering the test segments. read the letter →

arxiv 2601.10202 v2 pith:WF6XWF72 submitted 2026-01-15 q-bio.QM

Hongtao Li , Jia Wei , Jia Xiao , Yuanjun Lai , Mingyang Liu , Shuzhen Lv , Xueqiang Ouyang This is my paper
classification q-bio.QM
keywords atrialfibrillationdetectionelectrocardiogramtime-frequencyfusionsupervisedcontrastivelearningcross-datasetgeneralizationbidirectionalgatingdeepambulatoryECG
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 that atrial fibrillation detection from single-lead ECG can be simultaneously made more robust within a dataset and more transferable across datasets by fusing time- and frequency-domain information in a specific way: a bidirectional gating module lets the two modalities refine each other, and a supervised contrastive loss aligns same-label embeddings across modalities. If the claim holds, a single architecture can handle the noise and morphology variation seen in ambulatory recordings while still performing well when transferred to a different patient cohort and recording setup. On the two benchmark datasets, the proposed network reaches accuracies of 0.9878 and 0.9801 in five-fold patient-wise cross-validation, and cross-dataset accuracies of 0.9165 and 0.9507 in the two transfer directions, while staying within 1.96 million parameters and under one giga-FLOP. The ablation study attributes most of the cross-dataset gain to the two new modules, and the interpretability analysis shows the model attends to the P-wave region in AF segments, matching clinical knowledge.

What carries the argument

Bidirectional Gating Module (BGM): at each encoder level, a global average-pooled context vector from one modality is passed through a linear projection and sigmoid to produce channel-wise gates for the other modality; the gated features are downsampled and concatenated with the originals, enabling reciprocal refinement. Cross-modal Supervised Contrastive Learning (CSCL): after projecting the temporal and spectral embeddings to a shared D-dimensional space, the model applies two contrastive terms — intra-modal (same-label samples cluster within each modality) and inter-modal (same-label time-spectral pairs attract, different-label pairs repel) — so the joint embedding space is organized by c

What would settle it

Rerun the same training and cross-dataset evaluation without the quality-assessment filter, or evaluate the fitted model separately on the segments it discards; if accuracy on the excluded segments is near chance or the margin over baselines collapses, the robustness and generalization claims do not extend to the hardest real-world signals.

Watch

Extended reading notes

Core claim

The central claim is that the proposed MGCNet — a two-branch network feeding raw ECG through a 1D temporal encoder and an STFT spectrogram through a 2D spectral encoder — detects atrial fibrillation better than six recently published deep detectors, and does so under both same-dataset five-fold cross-validation and zero-fine-tuning cross-dataset transfer. The two named mechanisms carry the argument. The Bidirectional Gating Module uses channel-wise gating so that temporal features modulate spectral features and vice versa at three downsampling levels; the Cross-modal Supervised Contrastive Learning loss pulls same-class temporal and spectral embeddings together while pushing different-class

Load-bearing premise

The load-bearing premise is that the automated quality assessment that keeps only 'qualified' ECG segments does not systematically remove the noisy, difficult cases that real-world robustness claims are about; if it does, the reported gains are measured on a cleaned subsample, not on the full population of ambulatory recordings.

Editorial extensions

If this is right

  • AF detection on single-lead ECG can be made more robust to noise and morphological variation by explicit cross-modal gating rather than static feature concatenation or averaging.
  • Because the contrastive loss structures the joint embedding space by label, the same training recipe can improve zero-shot transfer between datasets with different sampling rates, device properties, and patient cohorts.
  • The reported efficiency (1.96 million parameters, 0.96 GFLOPs) means the contrastive-learning benefit comes with no extra inference cost, making the approach usable on portable or embedded monitors.
  • The ablation pattern — larger losses when BGM or CSCL is removed under transfer than within a dataset — indicates that both mechanisms contribute specifically to generalization, not just to memorization.
  • Grad-CAM attention on AF segments concentrates near the P-wave region, suggesting the learned representations align with the clinical criterion of absent or disorganized atrial activity.

Reading between the lines

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

  • The automated quality filter discards 'unqualified' segments before training and testing; if those are the noisy, difficult real-world cases, the reported robustness and transfer gains are measured on an enriched clean subset. An obvious check is to rerun the comparison without the filter, or report accuracy on the discarded segments.
  • The abstract promises bidirectional cross-dataset experiments involving LTAF and SHDBAF, but the results section only reports AFDB↔CPSC2021. Either the additional transfers are missing from the body or the scope was narrowed; a reader seeking evidence on broader generalization should look for those numbers.
  • The same two-module recipe — a bidirectional gating layer plus a supervised contrastive alignment of time and frequency embeddings — is not specific to AF or to the two tested datasets. A straightforward extension is to apply it to multi-class arrhythmia classification or to other biosignals where complementary time and spectral views carry distinct information.
  • The observed specificity-over-recall tradeoff under domain shift is a clinical design choice, but for a screening tool in a low-prevalence population the operating point should be evaluated with cost-sensitive metrics (e.g., net benefit or partial AUC), not only accuracy.
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.

Referee Report

4 major / 6 minor

Summary. This paper proposes MGCNet, a dual-branch deep network for AF detection from single-lead ECG. One branch encodes the raw time-domain signal with 1D convolutions and a BiGRU; the other encodes an STFT spectrogram with 2D convolutions. A Bidirectional Gating Module (BGM) performs channel-wise cross-modal gating, and a Cross-modal Supervised Contrastive Learning (CSCL) loss is added to the cross-entropy objective. The method is evaluated with patient-wise 5-fold cross-validation on AFDB and CPSC2021 and with bidirectional cross-dataset transfer between the two, comparing against six published baselines. The paper reports top accuracy/AUC on both datasets and both transfer directions (e.g., AFDB accuracy 0.9878, CPSC2021 accuracy 0.9801, AFDB→CPSC2021 accuracy 0.9165, CPSC2021→AFDB accuracy 0.9507), along with ablations, efficiency figures, and robustness to added interference.

Significance. If the reported results survive scrutiny, the contribution is valuable: a concrete, reproducible recipe combining time-frequency fusion with supervised contrastive learning, evaluated on public benchmarks under a patient-wise split and against several recent baselines. The ablation study and the public-data setup are strengths, and the claimed cross-dataset gains are clinically relevant. However, the headline generalization claim is currently undermined by two validation-protocol problems: (i) the contrastive weight λ is selected using one of the two cross-dataset target sets, and (ii) the quality filter discards unqualified segments from both training and test, so the reported numbers describe an enriched subset rather than the raw clinical distribution. These issues are fixable but require re-running experiments and additional reporting before the central claim can be accepted.

major comments (4)
  1. [Sensitivity Analysis / Implementation Details] The hyperparameter λ is selected on the target test set. In 'Sensitivity Analysis', λ ∈ {0,0.001,0.01,0.1,1.0} is varied and the paper states that performance 'peaked at λ=0.01' on CPSC2021 and under cross-dataset evaluation (CPSC2021→AFDB), after which 'Implementation Details' fixes λ=0.01 for all runs. Thus the reported CPSC2021→AFDB accuracy of 0.9507 (Table 4) is obtained with a hyperparameter chosen by looking at that exact target set. The AFDB→CPSC2021 result (0.9165) also uses a λ selected on the reverse transfer direction, so it is indirectly tuned as well. This is test-set hyperparameter tuning and biases the generalization claim. Fix: select λ using only source-domain validation data (or nested cross-validation), freeze it, and report cross-dataset results for that fixed λ across multiple seeds with mean and standard deviation.
  2. [Materials, Quality Assessment] The paper states 'Only qualified segments are retained for training and evaluation', applying an automated quality filter to both datasets. This removes noisy or artifact-contaminated segments from the test distribution before computing the reported accuracies. Since the paper's headline is robustness and generalization to real-world conditions, this filtering is load-bearing: the metrics are measured on an enriched, cleaned subset and may not reflect performance on the unqualified segments that are exactly the difficult cases. Please report the number and percentage of segments excluded for each dataset, and provide results on the full test set and/or on the unqualified subset. This applies equally to all compared methods, but it still limits the external-validity interpretation in the abstract and conclusions.
  3. [Table 4 / Generalization Experiments] Cross-dataset results are single runs with no variance estimates. Table 4 reports one accuracy number per method per direction, and no standard deviations, confidence intervals, or significance tests are given. Given stochastic training, data augmentation, and the small margins in some comparisons (e.g., 0.9165 vs 0.8951 for AFDB→CPSC2021; 0.9507 vs 0.9164 for CPSC2021→AFDB), these differences could be within seed-to-seed noise. Report at least 5 independent seeds with mean±std, and ideally a paired significance test over patients or folds.
  4. [Abstract / Conclusions] The abstract and conclusions claim 'consistent improvements over state-of-the-art methods across multiple metrics' and 'excellent cross-dataset generalization'. Table 4 shows that MGCNet has lower Recall than several baselines in both transfer directions (0.9007 vs up to 0.9579 in AFDB→CPSC2021; 0.9154 vs 0.9465 for MSCGN in CPSC2021→AFDB). The discussion acknowledges this as a specificity-precision trade-off, but the summary statements should be qualified so that the reader is not left with the impression of uniform superiority across all metrics.
minor comments (6)
  1. [Introduction / Abstract] Typos: 'majority existing methods' should be 'most existing methods'; 'robust and generalizable under' should be 'robust and generalizable to' or 'generalization'.
  2. [Comparative Experiments] Typo: 'compareed' should be 'compared'.
  3. [Figure 5] The figure axes are unlabeled in the manuscript text and the displayed values are hard to read. Add clear axis labels, tick labels, and ideally error bars for the sensitivity analysis.
  4. [Table 2] The 'Downsample' rows appear redundant with the 'BGM' rows that already halve the resolution. Clarify the data flow: is the downsampling performed inside BGM or as a separate operation, and what are the inputs to the 'Concat + Expand' step?
  5. [References] Reference [42] is formatted incorrectly ('Adam, K. D. B. J. et al.'); the correct citation is Kingma, D. P. and Ba, J., 'Adam: A Method for Stochastic Optimization.'
  6. [Data Availability] The sentence 'All code and trained models have be made publicly available' contains a typo ('have be' → 'have been'). Also, if the GitHub link is not yet active, state that it will be activated upon publication.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline cross-dataset generalization is partially circular: λ=0.01 was selected using the same target test sets whose results are later reported as predictions.

  1. fitted input called prediction [Sensitivity Analysis (Figure 5); Implementation Details; Generalization Experiments (Table 4)]
    "Performance peaked at λcont = 0.01, achieving an accuracy of 0.9801 and AUC of 0.9979 on CPSC2021, and 0.9507 / 0.9894 under cross-dataset evaluation. ... The contrastive loss hyperparameters were configured as λ=0.01 and τ=0.1."

    The final contrastive weight λ is selected after evaluating accuracy/AUC on CPSC2021 intra-dataset and on CPSC2021→AFDB cross-dataset. The same directions are then the headline cross-dataset results in Table 4 (AFDB→CPSC2021: 0.9165; CPSC2021→AFDB: 0.9507). Thus the reported 'without any fine-tuning or adaptation' generalization numbers are the best points of a small λ grid on the target test sets, not independent predictions. This is test-set hyperparameter selection on the central generalization claim, so the claimed cross-dataset advantage is partially circular/optimistically biased rather than a clean out-of-sample result.

full rationale

The only load-bearing circularity is the selection of λ=0.01 using the same target evaluations that are later presented as generalization results. The network design itself—BGM gating, CSCL losses, concatenation, classifier—is not derived from benchmark numbers; it is evaluated against them. No load-bearing self-citation or imported uniqueness theorem is present. The quality filter ('Only qualified segments are retained for training and evaluation') affects external validity by removing noisy segments, but it does not make the comparison circular. However, because both cross-dataset directions were used to tune a hyperparameter before their results were reported, the central generalization claim is partially reduced to a selected outcome. This is partial circularity in the evaluation protocol, not a definitional equivalence.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The method relies on standard deep-learning assumptions plus several domain choices: segment labeling, quality filtering, STFT representation, and the efficacy of supervised contrastive loss. No fundamentally new physical or mathematical entity is introduced; BGM and CSCL are architectural components validated only inside the paper. Hyperparameters, especially λ, are hand-tuned or tuned via sensitivity analysis on the same benchmarks used for the headline claims.

free parameters (6)
  • contrastive loss weight λ = 0.01
    Chosen by sensitivity analysis on CPSC2021 and CPSC2021→AFDB (Fig 5), i.e., tuned on a dataset that is later used to report the headline generalization result.
  • contrastive temperature τ = 0.1
    Hand-set; no sensitivity analysis reported.
  • STFT window and hop = 1-s Hann window, 50% overlap
    Fixed choices for the frequency-domain representation; not justified by comparison.
  • spectrogram cutoff/resolution = 40 Hz, resized to 128×128, RGB jet colormap
    Hand-set representation choices affecting frequency branch input; no ablation.
  • data augmentation ranges = noise σ∈[0,0.2], mask [0,0.1], amplitude U(0.5,1.5), zoom [0.8,1.2]
    Chosen by hand; no sensitivity analysis shown.
  • architecture channel widths = 32→64→128→256; embedding D unspecified
    Hand-designed dimensions listed in Table 2; no capacity ablation.
assumptions (5)
  • domain assumption 10-second non-overlapping segments inherit the entire recording's label
    Atrial fibrillation status can change within a recording; segment-level labels may mislabel brief sinus-rhythm periods inside AF episodes. This is the standard PhysioNet challenge setup and is assumed throughout Materials and Methods.
  • domain assumption Quality assessment correctly identifies unqualified segments and removing them improves generalization
    Section 'Quality Assessment' discards all unqualified segments before training/evaluation; if the scoring is imperfect, the reported accuracy is on a filtered population, not on all ambulatory ECGs.
  • domain assumption STFT magnitude spectrogram with frequencies ≤40 Hz contains sufficient discriminative information for AF
    Frequency branch is built from a 1-second Hann-window STFT clipped to 40 Hz; no phase information is kept, and this is asserted in 'Frequency-Domain Modality Generation' without comparison to other time-frequency representations.
  • domain assumption Supervised contrastive loss improves cross-dataset generalization in this setting
    The central design hypothesis behind CSCL; supported only by the paper's ablation, not by an external theoretical guarantee.
  • standard math Backpropagation and standard neural network training assumptions
    No formal verification; relies on standard optimization behavior.
invented entities (2)
  • BGM (Bidirectional Gating Module)
    purpose: Channel-wise gating so temporal and spectral feature maps modulate each other
    Only evaluated through internal ablation; no external dataset or theoretical result supports its contribution outside this paper's pipeline.
  • CSCL (Cross-modal Supervised Contrastive Learning)
    purpose: Supervised contrastive alignment of time and frequency embeddings
    Adapted from Khosla et al. (2020); the specific cross-modal form is new but is only internally ablated, not independently validated in another study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust and Generalizable Atrial Fibrillation Detection from ECG Using Time-Frequency Fusion and Supervised Contrastive Learning." pith.science (2026). https://pith.science/paper/WF6XWF72

@misc{pith2026260110202,
  author       = {Pith},
  title        = {Pith review of: Robust and Generalizable Atrial Fibrillation Detection from ECG Using Time-Frequency Fusion and Supervised Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WF6XWF72}},
  note         = {Machine review of arXiv:2601.10202}
}
read the original abstract

Atrial fibrillation (AF) is a common cardiac arrhythmia that significantly increases the risk of stroke and heart failure, necessitating reliable and generalizable detection methods from electrocardiogram (ECG) recordings. Although deep learning has advanced automated AF diagnosis, existing approaches often struggle to exploit complementary time frequency information effectively, limiting both robustness under intra-dataset and generalization across diverse clinical datasets. To address these challenges, we propose a crossmodal deep learning framework comprising two key components: a Bidirectional Gating Module (BGM) and a Cross modal Supervised Contrastive Learning (CSCL) strategy. The BGM facilitates dynamic, reciprocal refinement between time and frequency domain features, enhancing model robustness to signal variations within a dataset. Meanwhile, CSCL explicitly structures the joint embedding space by pulling together label consistent samples and pushing apart different ones, thereby improving interclass separability and enabling strong cross dataset generalization. We evaluate our method using five fold crossvalidation on the AFDB and CPSC2021 datasets. Furthermore, to assess cross dataset generalization, we conduct bidirectional cross dataset experiments across AFDB, CPSC2021, LTAF, and SHDBAF by training on one dataset and testing on another. Results show consistent improvements over state of the art methods across multiple metrics, demonstrating that our approach achieves both high intra dataset robustness and excellent crossdataset generalization. We further demonstrate that our method achieves high computational efficiency and anti interference capability, making it suitable for edge deployment.

Figures

Figures reproduced from arXiv: 2601.10202 by the authors.

Figure 1
Figure 1. Spectrogram of the cleaned ECG segment generated via STFT, encoded as a three-channel heatmap for frequency-domain modeling. In the time domain, AFIB is characterized by the absence of P waves and highly irregular R-R intervals. In the frequency domain (STFT spectrogram), it exhibits a diffuse and chaotic spectral pattern. All ECG recordings from the AFDB and the CPSC2021 are uniformly resampled to 250 Hz, and only … view at source ↗
Figure 2
Figure 2. (a) The full multimodal network for AF detection; (b) The BGM enabling dynamic interaction between time- and frequency-domain features; (c) The CSCL that enforces discriminative embedding alignment across modalities. 4/16 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the ECG data augmentation strategies applied in our experiments. (a) Original 10-second ECG segment; (b) with added Gaussian noise; (c) with random time masking; (d) with amplitude scaling; (e) with time zoom in; (f) with time zoom out. Implementation Details All experiments were implemented in PyTorch 2.4.1 with CUDA 11.5 acceleration and conducted on a workstation equipped with an NVIDIA GeForce RT… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Five-fold cross-validation results on two datasets across six metrics (Accuracy, AUC, F1-score, Precision, Recall and Specificity). (4) SeqAFNet19, a dual-path RNN framework designed to jointly model local and global temporal dependencies; (5) MFEGNet22, which employs …
Figure 5
Figure 5. Figure 5: Accuracy and AUC versus λcont on CPSC2021 (intra) and CPSC2021 → AFDB (cross). (a)Normal (b) Atrial fibrillation [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Grad-CAM attention maps overlaid on ECG waveforms. Redder regions indicate higher model attention, while bluer regions indicate lower attention. Top: Normal sinus rhythm shows diffuse activation across the cardiac cycle. Bottom: Atrial fibrillation exhibits concentrate…
Figure 7
Figure 7. Figure 7: Robustness of cross-dataset AF classification under increasing levels of ECG interference. Left column: performance of AFDB → CPSC2021 transfer; right column: CPSC2021 → AFDB. From top to bottom: (a,b) baseline drift(0.1 Hz, amplitude ratio 0–0.7), (c,d) powerline inte…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 17 canonical work pages

  1. [1]

    Cardiovasc

    Li, X.et al.Global burden and health inequality of atrial fibrillation/atrial flutter from 1990 to 2021.Front. Cardiovasc. Medicine12, 1585980, DOI: https://doi.org/10.3389/fcvm.2025.1585980 (2025)

    arXiv 1990

  2. [2]

    Bordignon, S., Corti, M. C. & Bilato, C. Atrial fibrillation associated with heart failure, stroke and mortality.J. Atr. Fibrillation5, 467, DOI: https://doi.org/10.4022/jafib.467 (2012)

    work page doi:10.4022/jafib.467 2012

  3. [3]

    Pamporis, K.et al.Prevalence of asymptomatic atrial fibrillation and risk factors associated with asymptomatic status: a systematic review and meta-analysis.Eur. J. Prev. Cardiol.zwaf138, DOI: https://doi.org/10.1093/eurjpc/zwaf138 (2025)

    work page doi:10.1093/eurjpc/zwaf138 2025

  4. [4]

    Y ., Huang, L., Malwade, S., Hsu, C.-Y

    Yang, T. Y ., Huang, L., Malwade, S., Hsu, C.-Y . & Chen, Y . C. Diagnostic accuracy of ambulatory devices in detecting atrial fibrillation: systematic review and meta-analysis.JMIR mHealth uHealth9, e26167, DOI: https://doi.org/10.2196/26167 (2021)

    work page doi:10.2196/26167 2021

  5. [5]

    Lip, G. Y . & Hee, F. L. S. Paroxysmal atrial fibrillation.Qjm94, 665–678, DOI: https://doi.org/10.1093/qjmed/94.12.665 (2001)

    work page doi:10.1093/qjmed/94.12.665 2001

  6. [6]

    1161/STROKEAHA.110.591958 (2010)

    Stahrenberg, R.et al.Enhanced detection of paroxysmal atrial fibrillation by early and prolonged continuous holter monitoring in patients with cerebral ischemia presenting in sinus rhythm.Stroke41, 2884–2888, DOI: https://doi.org/10. 1161/STROKEAHA.110.591958 (2010)

    2010

  7. [7]

    M.et al.Residual-attention deep learning model for atrial fibrillation detection from holter recordings.J

    Rahman, M. M.et al.Residual-attention deep learning model for atrial fibrillation detection from holter recordings.J. Electrocardiol.89, 153876, DOI: https://doi.org/10.1016/j.jelectrocard.2025.153876 (2025)

    arXiv 2025

  8. [8]

    Artificial intelligence with great potential in medical informatics: a brief review.Medinformatics1, 2–9, DOI: https://doi.org/10.47852/bonviewMEDIN42022204 (2024)

    Lin, H. Artificial intelligence with great potential in medical informatics: a brief review.Medinformatics1, 2–9, DOI: https://doi.org/10.47852/bonviewMEDIN42022204 (2024)

    work page doi:10.47852/bonviewmedin42022204 2024

Show all 44 references
  1. [10]

    Jabbar, A., Huang, J., Jabbar, M. K. & Ali, A. Adaptive multimodal fusion in vertical federated learning for decentralized glaucoma screening.Brain Sci.15, 990, DOI: https://doi.org/10.3390/brainsci15090990 (2025)

    2025 doi

  2. [11]

    R., Ulfenborg, B

    Stahlschmidt, S. R., Ulfenborg, B. & Synnergren, J. Multimodal deep learning for biomedical data fusion: a review. Briefings bioinformatics23, bbab569, DOI: https://doi.org/10.1093/bib/bbab569 (2022)

    2022 doi

  3. [12]

    journal environmental research public health18, 11302, DOI: https://doi.org/10.3390/ijerph182111302 (2021)

    Murat, F.et al.Review of deep learning-based atrial fibrillation detection studies.Int. journal environmental research public health18, 11302, DOI: https://doi.org/10.3390/ijerph182111302 (2021)

    2021 doi

  4. [13]

    & Akbilgic, O

    Mahajan, R., Kamaleswaran, R. & Akbilgic, O. Comparative analysis between convolutional neural network learned and engineered features: A case study on cardiac arrhythmia detection.Cardiovasc. Digit. Heal. J.1, 37–44, DOI: https://doi.org/10.1016/j.cvdhj.2020.04.001 (2020)

    2020 doi

  5. [14]

    Rao, S. S. & Martis, R. J. Rr interval-based atrial fibrillation detection using traditional and ensemble machine learning algorithms.J. Med. Signals & Sensors13, 224–232, DOI: https://doi.org/10.4103/jmss.jmss_4_22 (2023). 14/16

    2023 doi

  6. [15]

    & Zhang, H

    Xia, Y ., Wulan, N., Wang, K. & Zhang, H. Detecting atrial fibrillation by deep convolutional neural networks.Comput. biology medicine93, 84–92, DOI: https://doi.org/10.1016/j.compbiomed.2017.12.007 (2018)

    2017 doi

  7. [16]

    Faust, O.et al.Automated detection of atrial fibrillation using long short-term memory network with rr interval signals. Comput. biology medicine102, 327–335, DOI: https://doi.org/10.1016/j.compbiomed.2018.07.001 (2018)

    2018 doi

  8. [17]

    A.et al.Deep residual-dense network based on bidirectional recurrent neural network for atrial fibrillation detection.Sci

    Laghari, A. A.et al.Deep residual-dense network based on bidirectional recurrent neural network for atrial fibrillation detection.Sci. reports13, 15109, DOI: https://doi.org/10.1038/s41598-023-40343-x (2023)

    2023 doi

  9. [18]

    & Jang, J

    Kim, S., Lim, J. & Jang, J. Patient-adaptive beat-wise temporal transformer for atrial fibrillation classification in continuous long-term cardiac monitoring.IEEE AccessDOI: https://doi.org/10.1109/ACCESS.2024.3498043 (2024)

    2024

  10. [19]

    & Jang, J

    Kim, S., Lim, J. & Jang, J. Seqafnet: A beat-wise sequential neural network for atrial fibrillation classification in adhesive patch-type electrocardiographs.IEEE J. Biomed. Heal. Informatics28, 5260–5269, DOI: https://doi.org/10.1109/JBHI. 2024.3411056 (2024)

    2024

  11. [20]

    Signal Process

    Sun, Y .et al.Imc-resnet: Atrial fibrillation detection method based on interlayer multiscale coupling.Biomed. Signal Process. Control.97, 106683, DOI: https://doi.org/10.1016/j.bspc.2024.106683 (2024)

    2024

  12. [21]

    & Xie, L

    Wu, X., Yan, M., Tang, H., Wu, D. & Xie, L. Mscgn: Multiscale complementary gating network for time series classification. Biomed. Signal Process. Control.112, 108563, DOI: https://doi.org/10.1016/j.bspc.2025.108563 (2026)

    2025

  13. [22]

    & Xie, L

    Wu, X., Yan, M., Wang, R. & Xie, L. Multiscale feature enhanced gating network for atrial fibrillation detection.Comput. Methods Programs Biomed.261, 108606, DOI: https://doi.org/10.1016/j.cmpb.2025.108606 (2025)

    2025

  14. [23]

    Methods Programs Biomed.238, 107565, DOI: https://doi.org/10.1016/j.cmpb.2023.107565 (2023)

    Li, Y .et al.Diagnosis of atrial fibrillation using self-complementary attentional convolutional neural network.Comput. Methods Programs Biomed.238, 107565, DOI: https://doi.org/10.1016/j.cmpb.2023.107565 (2023)

    2023

  15. [24]

    physiology9, 1206, DOI: https://doi.org/10.3389/fphys.2018.01206 (2018)

    He, R.et al.Automatic detection of atrial fibrillation based on continuous wavelet transform and 2d convolutional neural networks.Front. physiology9, 1206, DOI: https://doi.org/10.3389/fphys.2018.01206 (2018)

    2018

  16. [25]

    & Zhang, Q

    Luo, Y ., Huang, B., Zhu, J., Zeng, X. & Zhang, Q. Detection of atrial fibrillation with a hybrid deep learning model and time-frequency representations.medRxiv2025–06, DOI: https://doi.org/10.1101/2025.06.03.25328929 (2025)

    2025 doi

  17. [26]

    & Zhou, X

    Yue, Y ., Chen, C., Liu, P., Xing, Y . & Zhou, X. Automatic detection of short-term atrial fibrillation segments based on frequency slice wavelet transform and machine learning techniques.Sensors21, 5302, DOI: https://doi.org/10.3390/ s21165302 (2021). 27.Zhu, H., Jiang, N., X...

    2021 doi

  18. [28]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, 770–778, DOI: https://doi.org/10.1109/CVPR.2016.90 (2016)

    2016 doi

  19. [29]

    & Yao, B

    Xie, J., Stavrakis, S. & Yao, B. Automated identification of atrial fibrillation from single-lead ecgs using multi-branching resnet.Front. Physiol.15, 1362185, DOI: https://doi.org/10.3389/fphys.2024.1362185 (2024)

    2024

  20. [30]

    informatics decision making20, 308, DOI: https://doi.org/10.1186/s12911-020-01337-1 (2020)

    Hu, Y .et al.An effective frequency-domain feature of atrial fibrillation based on time–frequency analysis.BMC Med. informatics decision making20, 308, DOI: https://doi.org/10.1186/s12911-020-01337-1 (2020)

    2020 doi

  21. [31]

    & Huang, W

    Ma, K., Zhang, T., Zhang, H. & Huang, W. Self-supervised contrastive learning achieves 12-lead ecg classification.Biomed. Signal Process. Control.112, 108420, DOI: https://doi.org/10.1016/j.bspc.2025.108420 (2026)

    2025

  22. [32]

    & Sakurai, T

    He, S., Ye, X. & Sakurai, T. Clop-herg: The contrastive learning optimized pre-trained model for representation learning in predicting drug-induced herg channel blockers.Medinformatics1, 103–111, DOI: https://doi.org/10.47852/ bonviewMEDIN42022049 (2024)

    2024

  23. [33]

    A new method for detecting atrial fibrillation using rr intervals.Proc

    Moody, G. A new method for detecting atrial fibrillation using rr intervals.Proc. Comput. Cardiol.10, 227–230, DOI: https://www.physionet.org/content/afdb/1.0.0/ (1983)

    1983

  24. [34]

    PhysioNet1–83, DOI: https://physionet.org/content/cpsc2021/1.0.0 (2021)

    Wang, X.et al.Paroxysmal atrial fibrillation events detection from dynamic ecg recordings: The 4th china physiological signal challenge 2021.Proc. PhysioNet1–83, DOI: https://physionet.org/content/cpsc2021/1.0.0 (2021)

    2021

  25. [35]

    K., Vijayalakshmi, D

    Anbalagan, T., Nath, M. K., Vijayalakshmi, D. & Anbalagan, A. Analysis of various techniques for ecg signal in healthcare, past, present, and future.Biomed. Eng. Adv.6, 100089, DOI: https://doi.org/10.1016/j.bea.2023.100089 (2023)

    2023

  26. [36]

    Attia, Z. I.et al.An artificial intelligence-enabled ecg algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction.The Lancet394, 861–867, DOI: https://doi.org/10. 1016/S0140-6736(19)31721-0 (2019). 15/16

    2019

  27. [37]

    & Zhang, Y

    Zhao, Z. & Zhang, Y . Sqi quality evaluation mechanism of single-lead ecg signal based on simple heuristic fusion and fuzzy comprehensive evaluation.Front. physiology9, 727, DOI: https://doi.org/10.3389/fphys.2018.00727 (2018)

    2018

  28. [38]

    research methods53, 1689–1696, DOI: https://github.com/neuropsychology/NeuroKit (2021)

    Makowski, D.et al.Neurokit2: A python toolbox for neurophysiological signal processing.Behav. research methods53, 1689–1696, DOI: https://github.com/neuropsychology/NeuroKit (2021). 39.Oppenheim, A. V .Discrete-time signal processing(Pearson Education India, 1999)

    2021

  29. [40]

    Cho, K.et al.Learning phrase representations using rnn encoder-decoder for statistical machine translation.arXiv preprint arXiv:1406.1078(https://doi.org/10.48550/arXiv.1406.1078)

    doi arXiv

  30. [41]

    neural information processing systems33, 18661–18673, DOI: https://doi.org/10.48550/arXiv.2004.11362 (2020)

    Khosla, P.et al.Supervised contrastive learning.Adv. neural information processing systems33, 18661–18673, DOI: https://doi.org/10.48550/arXiv.2004.11362 (2020)

    2004 doi arXiv

  31. [42]

    Adam, K. D. B. J.et al.A method for stochastic optimization.arXiv preprint arXiv:1412.69801412, DOI: https: //doi.org/10.48550/arXiv.1412.6980 (2014)

    2014 doi

  32. [43]

    R.et al.Grad-cam: Visual explanations from deep networks via gradient-based localization

    Selvaraju, R. R.et al.Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE international conference on computer vision, 618–626, DOI: https://doi.org/10.1109/ICCV .2017.74 (2017)

    2017 doi

  33. [44]

    Cinotti, E.et al.An edge ai approach for low-power, real-time atrial fibrillation detection on wearable devices based on heartbeat intervals.Sensors25, 7244, DOI: https://doi.org/10.3390/s25237244 (2025)

    2025 doi

  34. [45]

    Abdelrazik, A.et al.Wearable devices for arrhythmia detection: advancements and clinical implications.Sensors (Basel, Switzerland)25, 2848, DOI: https://doi.org/10.3390/s25092848 (2025)

    2025 doi

  35. [46]

    & Sunkaria, R

    Singh, O. & Sunkaria, R. K. Ecg signal denoising via empirical wavelet transform.Australas. physical & engineering sciences medicine40, 219–229, DOI: https://doi.org/10.1007/s13246-016-0510-6 (2017)

    2017 doi

  36. [47]

    & Sunkaria, R

    Singh, O. & Sunkaria, R. K. Powerline interference reduction in ecg signals using empirical wavelet transform and adaptive filtering.J. medical engineering & technology39, 60–68, DOI: https://doi.org/10.3109/03091902.2014.979954 (2015). 16/16

    2014

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.