Pith. sign in

REVIEW 5 major objections 6 minor 55 references

PhysioCLR integrates physiological similarity—feature-based pair selection, heartbeat shuffling, and peak-aware reconstruction—into ECG contrastive learning and reports consistent AUROC gains over stronger baselines across public and ICU da

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

2026-08-04 21:14 UTC pith:MH3AOELA

load-bearing objection PhysioCLR's pair-selection idea is fresh and the public-cohort results are plausible, but the private-cohort evaluation may be contaminated by finetuning on the test set, and the headline percentages don't match the tables. the 5 major comments →

arxiv 2509.08116 v1 pith:MH3AOELA submitted 2025-09-09 cs.LG cs.AI

Domain Knowledge is Power: Leveraging Physiological Priors for Self Supervised Representation Learning in Electrocardiography

Nooshin Maghsoodi , Sarah Nassar , Paul F R Wilson , Minh Nguyen Nhat To , Sophia Mannina , Shamel Addas , Stephanie Sibley , David Maslove
show 2 more authors
Purang Abolmaesumi Parvin Mousavi
This is my paper
classification cs.LG cs.AI
keywords self-supervised learningcontrastive learningelectrocardiogramarrhythmia classificationphysiological priorsrepresentation learningatrial fibrillationtransfer learning
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.

PhysioCLR argues that the key to better ECG self-supervised learning is not just more data or bigger models, but choosing which samples to compare. It builds positive and negative pairs from physiological similarity: ECG segments with close hand-crafted features (peak counts, amplitudes, RR intervals, heart-rate variability) are treated as similar, while dissimilar segments become negatives. It adds heartbeat shuffling as an augmentation that preserves intra-beat waveform morphology while disrupting inter-beat order, and a reconstruction loss that emphasizes waveform peaks. The paper reports that this physiology-aware recipe raises mean AUROC by 12% relative to the strongest baseline across Chapman, Georgia, and a noisy 4-lead ICU dataset, and that the gains hold when labeled fine-tuning data is scarce.

Core claim

The central claim is that physiological priors should be embedded in every design choice of ECG contrastive learning. Instead of relying only on patient-based temporal adjacency or generic augmentations, PhysioCLR derives positive and negative pairs from cosine similarity over PCA-reduced hand-crafted ECG features, generates an additional positive view by shuffling whole heartbeats, and trains a decoder with both global and peak-focused reconstruction losses. The paper demonstrates that these components together improve downstream arrhythmia classification: AUROC of 0.856 versus 0.821 on Chapman, 0.776 versus 0.729 on Georgia, and 0.922 versus 0.901 on the KGH ICU dataset, compared with the

What carries the argument

Feature-informed pair selection: up to 150 physiological features per ECG segment are extracted, zero-padded, normalized, PCA-reduced to 50 dimensions, and compared by cosine similarity against a threshold (delta=0.25 in the main runs); pairs above the threshold are positives, pairs below are negatives. Heartbeat shuffling: R-peaks segment the signal into heartbeats, which are randomly permuted and concatenated to form a positive view that keeps within-beat morphology intact. Peak-aware reconstruction: a decoder reconstructs the input, with a global mean-squared-error term and a peak-based term computed on detected local maxima. The total objective is L = L_contrastive + lambda * L_recon, wh

Load-bearing premise

The entire pretraining signal rests on hand-crafted ECG features—especially R-peak locations and derived intervals—being accurate enough that cosine similarity reflects true physiological similarity; on noisy ICU recordings where peak detection is unreliable, wrong positive and negative pairs would degrade the learned representations, and the paper itself notes that feature precision is critical.

What would settle it

Run the identical pretraining pipeline but replace feature-similarity pairs with random pairs, or with pairs generated from deliberately corrupted R-peak locations; if downstream AUROC on Chapman, Georgia, and KGH does not drop, the central claim that physiological similarity drives the gain is false. A second decisive test is to compare a version whose features are computed from clean, noise-free segments against one whose features are computed from noisy versions of the same segments.

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

If this is right

  • If physiological similarity, not patient identity, defines positives, the false-negative problem of assigning different patients with the same pathology as negatives is reduced, so embeddings should align with clinical meaning.
  • A model pretrained on large unlabeled 12-lead ECGs can transfer to a noisy 4-lead ICU setting, supporting lead-agnostic and environment-agnostic ECG analysis.
  • Under label scarcity, the pretrained model degrades less than supervised training: on Georgia, PhysioCLR drops from 0.776 to 0.741 when one labeled finetuning dataset is removed, while the supervised model drops from 0.724 to 0.667.
  • Each component contributes: feature-level pair selection alone yields a 2.49% average AUROC improvement over the strongest baseline, and the full combination yields a 4.39% average improvement.
  • The same design pattern—physiological similarity as pair supervision, structure-preserving augmentation, and peak-focused reconstruction—can be transferred to other biosignals with known event morphology.

Where Pith is reading between the lines

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

  • Beyond the paper: because the pretraining signal depends on hand-crafted features, a direct testable extension is to replace PCA/cosine-similarity selection with a learned similarity function; the paper's own conclusion flags feature precision as the critical vulnerability.
  • Beyond the paper: the optimal similarity threshold differs by dataset (0.25 for Chapman and Georgia, 0.5 for KGH), suggesting that a noise- or lead-count-adaptive threshold could further improve transfer.
  • Beyond the paper: the comparison to ECG-FM is only possible on KGH because ECG-FM was pretrained on the public test datasets, so the public-dataset gains should be read as gains over the W2V+CMSC+RLM line rather than over all foundation models.
  • Beyond the paper: a falsifiable prediction follows directly—corrupting or removing R-peak-dependent features during pair selection should degrade downstream AUROC; if it does not, the observed gains come from something other than physiological similarity.

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

5 major / 6 minor

Summary. The manuscript introduces PhysioCLR, a self-supervised ECG representation-learning method that integrates three physiology-informed components: feature-based positive/negative pair selection from hand-crafted physiological descriptors, heartbeat-shuffling augmentation, and a peak-aware reconstruction loss combined with a contrastive objective. The encoder follows the W2V+CMSC+RLM architecture of Oh et al., with a CNN frontend, transformer backbone, and random lead masking. Pretraining is performed on MIMIC-IV-ECG and PTB-XL; downstream evaluation is reported on Chapman, Georgia, and a private KGH ICU cohort. The paper claims a 12% relative mean AUROC improvement over the strongest baseline and ablation evidence that each proposed component contributes.

Significance. If the evaluation is clean, the paper makes a useful contribution to ECG self-supervised learning: it demonstrates a principled way to inject physiological priors into contrastive pair construction, and the use of a private ICU cohort is a meaningful generalization test. The comparison against a strong W2V+CMSC+RLM baseline and the large-scale pretraining setup are appropriate. The central idea is worth serious consideration. However, the reported quantitative claims are internally inconsistent, and the handling of the KGH dataset creates a risk that the headline cross-dataset generalization result is not an independent evaluation. These issues need to be resolved before the results can be relied upon.

major comments (5)
  1. The KGH evaluation protocol is ambiguous and potentially invalid. Table I designates KGH as a testing set, and Section IV.C describes evaluating the finetuned model on KGH, but Section IV.D states: 'Finetuning is performed on the labeled subset of PhysioNet 2021 and the KGH dataset.' If any KGH segments used for finetuning overlap with the 613 evaluation segments, the KGH row in Table II is not an independent generalization result, and the abstract's claim of robust cross-dataset generalization loses its private-cohort support. The paper must clarify whether KGH labels were used in finetuning, specify the exact KGH train/test split and any cross-validation, and if necessary re-run the evaluation with KGH excluded from training. The same clarification is needed for the phrase 'labeled subset of PhysioNet 2021' to confirm that Chapman and Georgia were not used in finetuning.
  2. The headline '12% relative improvement' is not supported by the numbers in Table II. Relative AUROC gains of PhysioCLR over the strongest baseline W2V+CMSC+RLM are: Chapman (0.856-0.821)/0.821 = 4.3%, Georgia (0.776-0.729)/0.729 = 6.5%, KGH (0.922-0.901)/0.901 = 2.3%. The mean relative gain is approximately 4.3%, not 12%. Please correct the abstract and any similar summary statements, or explicitly show the computation that yields 12%. This is not a cosmetic issue because the headline overstates the result by a factor of about three.
  3. The ablation results are numerically inconsistent. The text states that PhysioFeat alone yields a 2.49% average AUROC improvement, that the addition of ReconLoss yields an additional 9.5% improvement, and then that the full method achieves an average improvement of 4.39%. These numbers are mutually incompatible. Figure 6 does not display the per-component values. Please provide a corrected ablation table with per-dataset AUROC for each component (PhysioFeat, HRShuff, ReconLoss, and their combinations), as well as standard deviations or error bars. Without this, the central claim that all components contribute is not verifiable.
  4. The negative set definition is internally inconsistent. The text says negatives are 'all other elements in the batch that are not selected as positives,' but Eq. (1) defines N_i = {x_k | sim(z~_i, z~_k) < delta}. Since patient-based positives and heartbeat-shuffled positives are selected by temporal adjacency or shuffling, a positive sample can have cosine similarity below delta and thus be included in N_i. In Eq. (2), such a sample would then appear in both the numerator and denominator for the same anchor-positive pair. Please define N_i as batch\P_i, or explicitly exclude all elements of P_i from N_i, and adjust the loss accordingly. This is a technical flaw in the proposed objective that needs to be fixed.
  5. The use of CPSC is inconsistent across experiments. Section IV.A.1 states that CPSC and CPSC-Extra are used for validation and hyperparameter tuning, and Table I lists them under 'Validation (26-class multilabel classification).' However, Table III presents label-scarcity results with CPSC as one of the finetuning datasets. If CPSC labels were used to tune hyperparameters, then using CPSC also as a finetuning set in Table III is not a clean evaluation. Please clarify whether CPSC was held out from finetuning in the main experiments and, if it is used in the label-scarcity study, describe how hyperparameters were selected for that study and whether CPSC results generalize. This is load-bearing for the label-efficiency claim.
minor comments (6)
  1. Typo: 'ECG eatures' should be 'ECG features.'
  2. The statement 'Finetuning is performed on the labeled subset of PhysioNet 2021 and the KGH dataset' directly contradicts Table I and Section IV.A.1, which list finetuning as PTB-XL and Ningbo and KGH as testing. Please correct the wording or the experimental description.
  3. The sentence 'The addition of ReconLoss yields an additional 9.5% improvement' appears to be a typographical error, but as written it is impossible to reconcile with the 4.39% full-method improvement. Please correct the number and make the per-component increments explicit.
  4. The exact set of physiological features is described only as 'up to 150' and 'over 100.' Please provide a complete list of features extracted by NeuroKit2 and pyHRV, and state how the zero-padding to 150 dimensions is handled. Also, the PCA dimensionality (50) is a free parameter; consider reporting sensitivity to it.
  5. No confidence intervals or multiple-seed results are reported. Given the moderate differences against W2V+CMSC+RLM (e.g., KGH AUROC 0.922 vs 0.901), it would be helpful to report variance across seeds or at least state whether the reported numbers are single-run results.
  6. The threshold-sensitivity analysis is useful, but the text says KGH performance improves up to 0.5 while the implementation details use a threshold of 0.25 for the full model. Please clarify how the final threshold was chosen across datasets and whether a single threshold is used for all downstream tasks.

Circularity Check

1 steps flagged

KGH test-set contamination: the private-cohort AUROC is trained on the same data it is used to predict; public benchmark results remain independent.

specific steps
  1. fitted input called prediction [Section IV.D Implementation Details; Section IV.C Evaluation Protocol; Table I]
    "Finetuning is performed on the labeled subset of PhysioNet 2021 and the KGH dataset. ... Finally, we evaluate this model on the test sets. [Table I:] Testing (Binary classification): KGH Private Dataset 613"

    KGH is designated as a held-out test set in Table I and the Evaluation Protocol, but Section IV.D states that finetuning uses the KGH dataset. Therefore the reported KGH AUROC (0.922) is obtained by evaluating a model whose finetuning labels included the same KGH segments. This is not cross-dataset prediction; the private-cohort result is a fit to the test data by construction, and it inflates the abstract's claimed mean AUROC gain. No KGH train/test split or cross-validation is described, so as written the KGH evaluation is circular.

full rationale

The only clear circular step is the KGH evaluation. The paper is explicit that KGH is a testing dataset, yet the implementation details say finetuning is performed on the KGH dataset, making the KGH AUROC a trained-on-test result rather than a generalization measurement. This warrants a score of 6: one key 'prediction' reduces by construction, while the Chapman and Georgia results remain independent public-benchmark evidence. The physiological-feature-based pair selection is not circular: features are extracted from unlabeled signals and used only to define contrastive pairs, while downstream labels are separate; this is an intentional inductive bias, not a formal equivalence. The self-citations are not load-bearing, and the '12%' and ablation percentage inconsistencies are correctness/numerical issues, not circularity. The authors' own limitation about feature precision is a robustness concern, not a circularity concern.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on a feature-based similarity proxy whose components (RR intervals, HRV, peak morphology) are tuned/fitted and whose reliability is assumed; the encoder and data-split choices are adopted from prior work. No new entities (particles, forces, dimensions) are introduced.

free parameters (6)
  • cosine similarity threshold delta = 0.25
    Used in Eq. 1 to partition batch into positives/negatives; tuned to maximize validation AUROC; Fig. 5 shows strong sensitivity.
  • PCA output dimensionality = 50
    Features zero-padded to 150 then reduced to 50 via PCA; dimensionality chosen without stated justification.
  • reconstruction weight alpha = 0.2
    Weights global MSE in L_recon; tuned on validation.
  • reconstruction weight beta = 0.1
    Weights peak MSE in L_recon; tuned on validation.
  • hybrid loss weight lambda = not stated
    Appears in L_SSL = L_contrastive + lambda * L_recon; value not reported in the text.
  • peak prominence threshold = unspecified
    Used for peak detection in L_peaks; not specified in the text.
axioms (4)
  • domain assumption The hand-crafted physiological features (peak counts, amplitudes, RR intervals, HRV, slopes, energy) are reliable and capture clinically meaningful similarity.
    Invoked in Section III.B.2 for feature-based pair selection; on noisy ICU data the features may be corrupted.
  • ad hoc to paper Cosine similarity in PCA-reduced feature space is a valid proxy for physiological/clinical similarity of ECG segments.
    Defines the positive/negative sets in Eq. 1; no external validation of this proxy is provided.
  • domain assumption The Oh et al. encoder with random lead masking transfers across lead counts and datasets.
    Network architecture adopted from [20], used for 12-lead to 4-lead transfer; no analysis of lead-masking interaction with PhysioCLR components.
  • domain assumption Excluding PTB and INCART from downstream tasks because of long recordings is appropriate.
    Split selection rationale in Section IV.A.1; assumes global labels of long recordings are not suitable for 10-second segments.

pith-pipeline@v1.3.0-alltime-deepseek · 15378 in / 15154 out tokens · 148495 ms · 2026-08-04T21:14:44.455304+00:00 · methodology

0 comments
read the original abstract

Objective: Electrocardiograms (ECGs) play a crucial role in diagnosing heart conditions; however, the effectiveness of artificial intelligence (AI)-based ECG analysis is often hindered by the limited availability of labeled data. Self-supervised learning (SSL) can address this by leveraging large-scale unlabeled data. We introduce PhysioCLR (Physiology-aware Contrastive Learning Representation for ECG), a physiology-aware contrastive learning framework that incorporates domain-specific priors to enhance the generalizability and clinical relevance of ECG-based arrhythmia classification. Methods: During pretraining, PhysioCLR learns to bring together embeddings of samples that share similar clinically relevant features while pushing apart those that are dissimilar. Unlike existing methods, our method integrates ECG physiological similarity cues into contrastive learning, promoting the learning of clinically meaningful representations. Additionally, we introduce ECG- specific augmentations that preserve the ECG category post augmentation and propose a hybrid loss function to further refine the quality of learned representations. Results: We evaluate PhysioCLR on two public ECG datasets, Chapman and Georgia, for multilabel ECG diagnoses, as well as a private ICU dataset labeled for binary classification. Across the Chapman, Georgia, and private cohorts, PhysioCLR boosts the mean AUROC by 12% relative to the strongest baseline, underscoring its robust cross-dataset generalization. Conclusion: By embedding physiological knowledge into contrastive learning, PhysioCLR enables the model to learn clinically meaningful and transferable ECG eatures. Significance: PhysioCLR demonstrates the potential of physiology-informed SSL to offer a promising path toward more effective and label-efficient ECG diagnostics.

Figures

Figures reproduced from arXiv: 2509.08116 by David Maslove, Minh Nguyen Nhat To, Nooshin Maghsoodi, Parvin Mousavi, Paul F R Wilson, Purang Abolmaesumi, Sarah Nassar, Shamel Addas, Sophia Mannina, Stephanie Sibley.

Figure 1
Figure 1. Figure 1: Overview of the proposed model: (a) Each ECG segment, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Feature-Level Sample Selection: This diagram shows how contrastive pairs are created using ECG-specific features for self-supervised learning. Each [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Heartbeat Shuffling Augmentation. In this augmentation technique, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The Encoder Architecture is inspired by the model proposed by [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of similarity threshold on AUROC performance metric for [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation Study on Proposed Method Components: Each group [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] 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

55 extracted references · 46 canonical work pages · 1 internal anchor

  1. [1]

    Deep learning for healthcare applications based on physiological signals: A review,

    O. Faust, Y . Hagiwara, T. J. Hong, R. S. Tan, and U. R. Acharya, “Deep learning for healthcare applications based on physiological signals: A review,”Comput. Methods Programs Biomed., vol. 161, pp. 1–13, 2018

    work page 2018

  2. [2]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov et al., “DINOv2: Learning robust visual features without supervision,” arXiv:2304.07193, Apr. 2023

    Pith/arXiv arXiv 2023

  3. [3]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi and G. Hinton, “A simple framework for contrastive learning of visual representations,” inProc. 37th Int. Conf. Machine Learning (ICML), Nov. 2020, pp. 1597–1607

    work page 2020

  4. [4]

    Bootstrap your own latent: A new approach to self-supervised learning,

    J.-B. Grill, F. Strub, F. Altch ´eet al., “Bootstrap your own latent: A new approach to self-supervised learning,”Adv. Neural Inf. Process. Syst., vol. 33, pp. 21271–21284, 2020

    work page 2020

  5. [5]

    Context autoencoder for self- supervised representation learning,

    X. Chen, M. Ding, X. Wanget al., “Context autoencoder for self- supervised representation learning,”Int. J. Comput. Vis., vol. 132, no. 1, pp. 208–223, Jan. 2024, doi: 10.1007/s11263-023-01852-4

    doi:10.1007/s11263-023-01852-4 2024

  6. [6]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv:1807.03748, Jul. 2018

    Pith/arXiv arXiv 2018

  7. [7]

    Masked autoencoders for point-cloud self-supervised learning,

    Y . Pang, W. Wang, F. E. Tay, W. Liu, Y . Tian and L. Yuan, “Masked autoencoders for point-cloud self-supervised learning,” inProc. ECCV, Cham, Switzerland: Springer, Oct. 2022, pp. 604–621

    work page 2022

  8. [8]

    Self-supervised speech representation learning: A review,

    A. Mohamed, H.-Y . Lee, L. Borgholtet al., “Self-supervised speech representation learning: A review,”IEEE J. Sel. Top. Signal Process., vol. 16, no. 6, pp. 1179–1210, Dec. 2022

    work page 2022

  9. [9]

    Which augmentation should I use? An empirical investigation of augmentations for self-supervised phonocardiogram representation learning,

    A. R. Alkhulaifi et al., “Which augmentation should I use? An empirical investigation of augmentations for self-supervised phonocardiogram representation learning,” inProc. Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 1–4, July 2023

    work page 2023

  10. [10]

    Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,

    A. Y . Hannun, P. Rajpurkar, M. Haghpanahiet al., “Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,”Nat. Med., vol. 25, no. 1, pp. 65–69, Jan. 2019; see also published correction, doi: 10.1038/s41591-019-0359-9

    work page doi:10.1038/s41591-019-0359-9 2019

  11. [11]

    AF classification from a short single-lead ECG recording: The PhysioNet/Computing in Cardiology Challenge 2017,

    G. D. Clifford, C. Liu, B. Moodyet al., “AF classification from a short single-lead ECG recording: The PhysioNet/Computing in Cardiology Challenge 2017,”Comput. Cardiol., vol. 44, pp. 1–4, Sep. 2017

    work page 2017

  12. [12]

    Deep convolutional neural network for the automated diagnosis of congestive heart failure using ECG signals,

    U. R. Acharya, H. Fujita, S. L. Ohet al., “Deep convolutional neural network for the automated diagnosis of congestive heart failure using ECG signals,”Appl. Intell., vol. 49, no. 1, pp. 16–27, 2019

    work page 2019

  13. [13]

    ECG segmentation using a deep learning model,

    N. Ibtehaz, M. H. Mahmud and A. B. M. Al Islam, “ECG segmentation using a deep learning model,”Biocybern. Biomed. Eng., vol. 42, no. 2, pp. 418–431, 2022

    work page 2022

  14. [14]

    Self-supervised representation learning from 12-lead ECG data,

    T. Mehari and N. Strodthoff, “Self-supervised representation learning from 12-lead ECG data,”Comput. Biol. Med., vol. 141, Art. no. 105114, Feb. 2022

    work page 2022

  15. [15]

    3KG: Contrastive learning of 12-lead electrocardiograms using physiologically-inspired augmentations,

    B. Gopal, R. Han, G. Raghupathi, A. Ng, G. Tison and P. Ra- jpurkar, “3KG: Contrastive learning of 12-lead electrocardiograms using physiologically-inspired augmentations,” inProc. Machine Learning for Health (ML4H), PMLR vol. 158, pp. 156–167, Dec. 2021

    work page 2021

  16. [16]

    sCL-ST: Su- pervised contrastive learning with semantic transformations for multiple- lead ECG arrhythmia classification,

    D. Le, S. Truong, P. Brijesh, D. A. Adjeroh and N. Le, “sCL-ST: Su- pervised contrastive learning with semantic transformations for multiple- lead ECG arrhythmia classification,”IEEE J. Biomed. Health Inform., vol. 27, no. 6, pp. 2818–2828, Jun. 2023

    work page 2023

  17. [17]

    CLECG: A novel contrastive learning framework for electrocardiogram arrhythmia classification,

    H. Chen, G. Wang, G. Zhang, P. Zhang and H. Yang, “CLECG: A novel contrastive learning framework for electrocardiogram arrhythmia classification,”IEEE Signal Process. Lett., vol. 28, pp. 1993–1997, Dec. 2021

    work page 1993

  18. [18]

    CLOCS: Contrastive learning of cardiac signals across space, time and patients,

    D. Kiyasseh, T. Zhu and D. A. Clifton, “CLOCS: Contrastive learning of cardiac signals across space, time and patients,” inProc. 38th ICML, PMLR vol. 139, pp. 5606–5615, Jul. 2021

    work page 2021

  19. [19]

    Contrast everything: A hierarchical contrastive framework for medical time-series,

    Y . Wang, Y . Han, H. Wang and X. Zhang, “Contrast everything: A hierarchical contrastive framework for medical time-series,”Adv. Neural Inf. Process. Syst., vol. 36, Art. no. 15548, Dec. 2024

    work page 2024

  20. [20]

    Lead-agnostic self-supervised learning for local and global representations of electrocardiogram,

    J. Oh, Y . Lee and J. Kim, “Lead-agnostic self-supervised learning for local and global representations of electrocardiogram,” inProc. Machine Learning for Health (ML4H), PMLR vol. 174, pp. 322–337, Dec. 2022

    work page 2022

  21. [21]

    Multi-channel masked autoen- coder and comprehensive evaluations for reconstructing 12-lead ECG from arbitrary single-lead ECG,

    J. Chen, W. Wu, T. Liu and S. Hong, “Multi-channel masked autoen- coder and comprehensive evaluations for reconstructing 12-lead ECG from arbitrary single-lead ECG,”npj Cardiovasc. Health, vol. 1, no. 1, Art. no. 13, 2024

    work page 2024

  22. [22]

    Guiding masked representation learning to capture spatio-temporal relationships of electrocardiogram,

    Y . Na, M. Park, Y . Tae and S. Joo, “Guiding masked representation learning to capture spatio-temporal relationships of electrocardiogram,” arXiv:2402.09450, Feb. 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Learn- ing representations for multi-lead electrocardiograms from morphol- ogy–rhythm contrast,

    W. Liu, H. Zhang, S. Chang, H. Wang, J. He and Q. Huang, “Learn- ing representations for multi-lead electrocardiograms from morphol- ogy–rhythm contrast,”IEEE Trans. Instrum. Meas., early access, pp. 1–12, Jan. 2025, doi: 10.1109/TIM.2025.3274458

    work page arXiv 2025

  24. [24]

    Self-supervised inter–intra period-aware ECG representation learning for detecting atrial fibrillation,

    X. Zhou, M. Shi, X. Yuet al., “Self-supervised inter–intra period-aware ECG representation learning for detecting atrial fibrillation,”Biomed. Signal Process. Control, vol. 100, Art. no. 106939, 2025

    work page 2025

  25. [25]

    G. D. Clifford, F. Azuaje and P. E. McSharry,Advanced Methods and Tools for ECG Data Analysis. Norwood, MA, USA: Artech House, 2006

    work page 2006

  26. [26]

    Automated identification of shockable and non-shockable life-threatening ventricular arrhythmias using convolutional neural network,

    U. R. Acharya, H. Fujita, S. L. Ohet al., “Automated identification of shockable and non-shockable life-threatening ventricular arrhythmias using convolutional neural network,”Future Gener. Comput. Syst., vol. 79, pp. 952–959, 2018

    work page 2018

  27. [27]

    Surawicz and T

    B. Surawicz and T. Knilans,Chou’s Electrocardiography in Clinical Practice. 6th ed. Amsterdam, The Netherlands: Elsevier, 2008

    work page 2008

  28. [28]

    Support vector machine- based expert system for reliable heartbeat recognition,

    S. Osowski, L. T. Hoai and T. Markiewicz, “Support vector machine- based expert system for reliable heartbeat recognition,”IEEE Trans. Biomed. Eng., vol. 51, no. 4, pp. 582–589, Apr. 2004

    work page 2004

  29. [29]

    Automated patient-specific clas- sification of premature ventricular contractions,

    T. Ince, S. Kiranyaz and M. Gabbouj, “Automated patient-specific clas- sification of premature ventricular contractions,” inProc. 30th Annu. Int. Conf. IEEE EMBC, Vancouver, BC, Canada, Aug. 2008, pp. 5474–5477

    work page 2008

  30. [30]

    ECG feature extraction and classification using wavelet transform and support vector machines,

    Q. Zhao and L. Zhang, “ECG feature extraction and classification using wavelet transform and support vector machines,” inProc. Int. Conf. Neural Networks and Brain, Beijing, China, Oct. 2005, vol. 2, pp. 1089–1092

    work page 2005

  31. [31]

    Patient-specific ECG classification by deeper CNN from generic to dedicated,

    Y . Li, Y . Pang, J. Wang and X. Li, “Patient-specific ECG classification by deeper CNN from generic to dedicated,”Neurocomputing, vol. 314, pp. 336–346, Nov. 2018

    work page 2018

  32. [32]

    A novel imbalanced-dataset mitigation and ECG classification model based on combined 1D CBAM autoen- coder and lightweight CNN,

    Z. Chen, D. Yang, T. Cuiet al., “A novel imbalanced-dataset mitigation and ECG classification model based on combined 1D CBAM autoen- coder and lightweight CNN,”Biomed. Signal Process. Control, vol. 87, Art. no. 105437, 2024. 11

    work page 2024

  33. [33]

    A deep learning model for the classification of atrial fibrillation in critically ill patients,

    B. Chen, D. M. Maslove, J. D. Curranet al., “A deep learning model for the classification of atrial fibrillation in critically ill patients,”Intensive Care Med. Exp., vol. 11, Art. no. 2, Jan. 2023

    work page 2023

  34. [34]

    Classification of ECG arrhythmia using recurrent neural networks,

    S. Singh, S. K. Pandey, U. Pawar and R. R. Janghel, “Classification of ECG arrhythmia using recurrent neural networks,”Procedia Comput. Sci., vol. 132, pp. 1290–1297, 2018

    work page 2018

  35. [35]

    Deep learning-based classification of ECG signals using RNN and LSTM mechanism,

    V . Satheeswaran, G. N. Chandrika, A. Mitraet al., “Deep learning-based classification of ECG signals using RNN and LSTM mechanism,”J. Electron. Electromed. Eng. Med. Inform., vol. 6, no. 4, pp. 332–342, 2024

    work page 2024

  36. [36]

    Generative adversarial network with transformer generator for boosting ECG classification,

    Y . Xia, Y . Xu, P. Chen, J. Zhang and Y . Zhang, “Generative adversarial network with transformer generator for boosting ECG classification,” Biomed. Signal Process. Control, vol. 80, Art. no. 104276, 2023

    work page 2023

  37. [37]

    ECGTransForm: Empowering adap- tive ECG arrhythmia classification framework with bidirectional trans- former,

    H. El-Ghaish and E. Eldele, “ECGTransForm: Empowering adap- tive ECG arrhythmia classification framework with bidirectional trans- former,”Biomed. Signal Process. Control, vol. 89, Art. no. 105714, 2024

    work page 2024

  38. [38]

    ECG-FM: An open electrocar- diogram foundation model,

    K. McKeen, L. Oliva, S. Masoodet al., “ECG-FM: An open electrocar- diogram foundation model,”arXiv:2408.05178, May 2025

    Pith/arXiv arXiv 2025

  39. [39]

    DinoSR: Self- distillation and online clustering for self-supervised speech representa- tion learning,

    A. H. Liu, H.-J. Chang, M. Auli, W.-N. Hsu and J. Glass, “DinoSR: Self- distillation and online clustering for self-supervised speech representa- tion learning,”Adv. Neural Inf. Process. Syst., vol. 36, pp. 58346–58362, 2023

    work page 2023

  40. [40]

    Adversarial spatiotemporal contrastive learning for electrocardiogram signals,

    N. Wang, P. Feng, Z. Ge, Y . Zhou, B. Zhou and Z. Wang, “Adversarial spatiotemporal contrastive learning for electrocardiogram signals,”IEEE Trans. Neural Netw. Learn. Syst., vol. 35, no. 10, pp. 13845–13859, Oct. 2024

    work page 2024

  41. [41]

    Boosting contrastive self-supervised learning with false-negative can- cellation,

    T. Huynh, S. Kornblith, M. R. Walter, M. Maire and M. Khademi, “Boosting contrastive self-supervised learning with false-negative can- cellation,” inProc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), Jan. 2022, pp. 2785–2795

    work page 2022

  42. [42]

    MaeFE: Masked autoencoders family of electrocardiogram for self-supervised pre-training and transfer learning,

    H. Zhang, W. Liu, J. Shiet al., “MaeFE: Masked autoencoders family of electrocardiogram for self-supervised pre-training and transfer learning,” IEEE Trans. Instrum. Meas., vol. 72, pp. 1–16, 2023

    work page 2023

  43. [43]

    Masked Transformer for Electrocardiogram Classification

    Y . Zhou, X. Diao, Y . Huoet al., “Masked transformer for electrocardio- gram classification,”arXiv:2309.07136, Sep. 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  44. [44]

    RE- BAR: Retrieval-based reconstruction for time-series contrastive learn- ing,

    M. A. Xu, A. Moreno, H. Wei, B. M. Marlin and J. M. Rehg, “RE- BAR: Retrieval-based reconstruction for time-series contrastive learn- ing,”arXiv:2311.00519, Nov. 2023

    Pith/arXiv arXiv 2023

  45. [45]

    C-MELT: Contrastive enhanced masked auto-encoders for ECG-language pre-training,

    M. Pham, A. Saeed and D. Ma, “C-MELT: Contrastive enhanced masked auto-encoders for ECG-language pre-training,”arXiv:2410.02131, Oct. 2024

    Pith/arXiv arXiv 2024

  46. [46]

    NeuroKit2: A python toolbox for neurophysiological signal processing,

    D. Makowski, T. Pham, Z. J. Lauet al., “NeuroKit2: A python toolbox for neurophysiological signal processing,”Behav. Res. Methods, vol. 53, no. 4, pp. 1689–1696, Aug. 2021; erratum, doi: 10.1038/s41597-022- 01643-5

    doi:10.1038/s41597-022- 2021

  47. [47]

    A real-time QRS detection algorithm,

    J. Pan and W. J. Tompkins, “A real-time QRS detection algorithm,”IEEE Trans. Biomed. Eng., vol. 32, no. 3, pp. 230–236, Mar. 1985

    work page 1985

  48. [48]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, H. Zhou, A. Mohamed and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” arXiv:2006.11477, Jun. 2020

    Pith/arXiv arXiv 2006

  49. [49]

    PhysioBank, Phys- ioToolkit and PhysioNet: Components of a new research resource for complex physiologic signals,

    A. L. Goldberger, L. A. Amaral, L. Glasset al., “PhysioBank, Phys- ioToolkit and PhysioNet: Components of a new research resource for complex physiologic signals,”Circulation, vol. 101, no. 23, pp. e215–e220, Jun. 2000

    work page 2000

  50. [50]

    MIMIC-IV-ECG: Diagnostic electrocardiogram matched subset (version 1.0),

    B. Gow, T. Pollard, L. A. Nathansonet al., “MIMIC-IV-ECG: Diagnostic electrocardiogram matched subset (version 1.0),”PhysioNet, 2023

    work page 2023

  51. [51]

    Will two do? Varying dimensions in electrocardiography: The PhysioNet/Computing in Cardi- ology Challenge 2021,

    M. A. Reyna, N. Sadr, E. A. P. Aldayet al., “Will two do? Varying dimensions in electrocardiography: The PhysioNet/Computing in Cardi- ology Challenge 2021,” inProc. Comput. Cardiol. (CinC), Brno, Czech Republic, Sep. 2021, pp. 1–4

    work page 2021

  52. [52]

    Classification of 12- lead ECGs: The PhysioNet/Computing in Cardiology Challenge 2020,

    E. A. Perez Alday, A. Gu, A. J. Shahet al., “Classification of 12- lead ECGs: The PhysioNet/Computing in Cardiology Challenge 2020,” Physiol. Meas., vol. 41, no. 12, Art. no. 124003, Jan. 2021

    work page 2020

  53. [53]

    fairseq: A fast, extensible toolkit for sequence modeling,

    M. Ott, S. Edunov, A. Baevskiet al., “fairseq: A fast, extensible toolkit for sequence modeling,” inProc. NAACL-HLT 2019 (Demonstrations), Minneapolis, MN, USA, Jun. 2019, pp. 48–53

    work page 2019

  54. [54]

    fairseq-signals: Self-supervised learning framework for biosig- nals (ECG, PPG),

    J. Oh, “fairseq-signals: Self-supervised learning framework for biosig- nals (ECG, PPG),” GitHub repository, https://github.com/Jwoo5/fairseq- signals, accessed May 21, 2025

    work page 2025

  55. [55]

    What makes for good views for contrastive learning?

    Y . Tian, C. Sun, B. Poole, D. Krishnan, C. Schmid and P. Isola, “What makes for good views for contrastive learning?”Adv. Neural Inf. Process. Syst., vol. 33, pp. 6827–6839, 2020

    work page 2020