Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Self-DANA: A Resource-Efficient Channel-Adaptive Self-Supervised Approach for ECG Foundation Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Self-DANA, a channel-adaptive pooling layer combined with a new random-lead-selection augmentation, lets a single ECG foundation model fine-tune on any reduced-lead configuration with accuracy comparable to zero-padding while cutting…

desk verdict The DAP layer's resource-efficiency win is real and well-demonstrated; the RLS augmentation's performance edge over RLM is not statistically supported. read the letter →

arxiv 2507.14151 v1 pith:ARPSRJXG submitted 2025-07-03 eess.SP cs.AIcs.LG

classification eess.SPcs.AIcs.LG
keywords ECGfoundationmodelself-supervisedlearningchannel-adaptivereduced-leaddimensionadaptivepoolingrandomleadselectioncontrastiveresourceefficiency
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

Self-DANA is a method for making ECG foundation models work with any number of input leads, from the full 12-lead clinical setup down to the single lead of a smartwatch. It claims that one model pre-trained on 12-lead data can be fine-tuned on a reduced-lead configuration without the standard trick of zero-padding the missing leads, because two simple ingredients replace that trick: a dimension-adaptive pooling (DAP) layer that folds whatever leads are present into a fixed-size representation, and a new Random Lead Selection (RLS) augmentation that teaches the model during pre-training to make sense of arbitrary lead subsets. On an out-of-domain 23-class cardiac abnormality task, the paper reports that this combination matches or beats the zero-padding baseline on all five tested lead configurations while using up to 69.3% less peak CPU memory, 34.4% less peak GPU memory, about 17% less average epoch CPU time, and about 24% less average epoch GPU time. A sympathetic reader would care because this is a step toward one reusable ECG model that can be deployed efficiently on wearables and portable devices without retraining per device.

What carries the argument

The load-bearing mechanism is the pair formed by the Dimension Adaptive Pooling (DAP) layer and the Random Lead Selection (RLS) augmentation. DAP is an adaptive average-pooling layer placed between the convolutional feature encoder and the transformer encoder; it maps any feature map with $C$ channels and 156 temporal positions to a fixed $(1, 156)$ map by averaging over the channel dimension, so the transformer never sees the channel count. Because the preceding convolutions use kernels of shape $(1, k)$ along time only, each lead is convolved independently, which keeps the representation channel-count-independent. RLS is a contrastive augmentation that, for each positive pair during SimCLR pre-training, randomly chooses both the number (between 1 and 12) and the identity of the leads to keep, so the model learns to produce similar representations from different subsets of the same recording. Together they remove the need to zero-pad missing leads during fine-tuning, which is what saves memory and time, and they also make the pre-trained features robust to which leads a wearable happens to provide.

What would settle it

On a downstream task where the abnormality is visible in only one or two specific leads, for example, high-lateral ST-elevation infarction best seen in lead aVL or a posterior infarction best seen in leads V7-V9, fine-tune Self-DANA and the zero-padded RLM baseline on identical data and compare class-wise sensitivity; if Self-DANA's sensitivity on those lead-specific classes drops by more than a few points relative to the baseline, the channel-averaging assumption is falsified for that task.

Watch

Extended reading notes

Core claim

The paper establishes that a channel-adaptive ECG foundation model, which it calls Self-DANA, can replace zero-padding as the way to handle reduced-lead downstream tasks. Its architecture uses four temporal-only 2D convolutions followed by a Dimension Adaptive Pooling layer that average-pools the channel dimension of the feature map, reducing any feature map of shape $(C, 156)$ to the fixed $(1, 156)$ token sequence the transformer encoder expects. During SimCLR pre-training, a new augmentation called Random Lead Selection randomly retains between 1 and 12 of the 12 leads for each positive pair, training the backbone to build consistent representations from arbitrary lead combinations. Fine-tuning then feeds only the channels actually available. Across 12-, 6-, 3-, 2-, and 1-lead configurations on the Georgia dataset, Self-DANA achieves CinC scores of $0.619 \pm 0.009$, $0.613 \pm 0.004$, $0.617 \pm 0.009$, $0.618 \pm 0.008$, and $0.583 \pm 0.004$ respectively, compared with $0.612 \pm 0.006$ through $0.585 \pm 0.004$ for the zero-padded RLM baseline, while consuming up to 69.3% less peak CPU memory and 34.4% less peak GPU memory during fine-tuning.

Load-bearing premise

Average pooling over the available leads preserves the diagnostic information carried by the specific lead combination, so a model pre-trained on random lead selections transfers to any fixed reduced-lead set without losing clinically meaningful details.

Editorial extensions

If this is right

  • One ECG foundation model pre-trained on 12-lead data can be fine-tuned on any reduced-lead wearable configuration without retraining a channel-specific model, since DAP removes the need to pad missing leads.
  • Fine-tuning memory now scales with the number of leads actually available: peak GPU memory falls from 23.19 GB to 15.21 GB between the padded and DAP versions at a single lead, enabling training on far more memory-constrained hardware.
  • The RLS augmentation improves the CinC score even in the 12-lead configuration (0.619 vs 0.612), so it acts as a general-purpose ECG augmentation rather than only a channel-adaptation trick.
  • Because the resource savings grow as leads are removed (up to 69.3% CPU memory and 34.4% GPU memory), the method's value increases precisely in the single-lead smartwatch and patch-monitor scenarios where wearables operate.
  • Fine-tuning a channel-adaptive foundation model beats training dedicated supervised models from scratch on each lead configuration, reinforcing the foundation-model transferability story.

Reading between the lines

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

  • Extending the DAP layer to learned per-lead attention weighting, trained under the same RLS objective, could preserve lead-specific morphology that plain averaging blurs while keeping the channel-count flexibility and most of the resource savings.
  • RLS could also be used as test-time augmentation: for a device with a fixed reduced-lead set, sampling different subsets of the available leads during inference might stabilize predictions without any additional fine-tuning.
  • The pairing of DAP with random channel selection is not ECG-specific; the same recipe could let one EEG or wearable-PPG model serve devices with 4-32 channels or varying sensor counts, where device heterogeneity currently forces per-configuration models.
  • The efficiency gains were measured on one GPU model and one batch size; re-running the same comparison at larger batch sizes or on edge hardware would show whether the memory savings compound or shrink in those deployment regimes.
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

3 major / 5 minor

Summary. The paper proposes Self-DANA, a channel-adaptive approach for ECG foundation models, combining a Dimension Adaptive Pooling (DAP) layer, which replaces zero-padding by pooling over the available channel dimension, with a new contrastive augmentation called Random Lead Selection (RLS), which feeds random subsets of leads during SimCLR pre-training. The authors pre-train on seven public 12-lead ECG datasets, fine-tune on the Georgia dataset in five reduced-lead configurations (12, 6, 3, 2, and 1 lead), and evaluate both CinC diagnostic performance and CPU/GPU memory and time. They report that DAP matches zero-padding in performance, that RLS improves over the base augmentations, that Self-DANA is competitive with or better than the prior RLM+zero-padding approach, and that it reduces peak CPU memory by up to 69.3% and peak GPU memory by up to 34.4%.

Significance. If the performance claims hold, this is a practically useful contribution: it shows a route to a single ECG foundation model that can be fine-tuned efficiently on arbitrary reduced-lead configurations, which is relevant for wearable and portable ECG devices. The strengths of the paper are its well-structured ablation design (FT-base-ZP vs FT-base-DAP isolates the DAP effect; Self-DANA vs FT-base-DAP isolates the RLS effect; multiple seeds are used), the out-of-domain evaluation on Georgia, the detailed architectural and dataset appendix, and the extended single-lead analysis in Table A6. The resource-efficiency results are substantial and clearly presented. However, the performance comparisons lack statistical significance testing, and one of the key interpretive claims about RLS being 'customized for DAP' rests on a comparison that is not fully controlled.

major comments (3)
  1. [§4.1(ii), Table 2] The central performance claim that Self-DANA reaches state-of-the-art or at least comparable performance is not supported by statistical significance testing. With only five seeds, several pairwise differences are within one standard deviation of the means, and the 1-lead comparison is numerically reversed (Self-DANA 0.583 ± 0.004 vs FT-RLM-ZP 0.585 ± 0.004). The statements that Self-DANA gives 'consistently higher performance' over FT-base-DAP and is 'comparable to, if not slightly higher than, FT-RLM-ZP' should be backed by paired significance tests across the five seeds, or by confidence intervals or effect sizes. Without these, the abstract's 'reaching state-of-the-art performance' and 'comparable or better' wording is too strong.
  2. [§3.5(ii) and Table 2] The claim that RLS is an ad-hoc augmentation 'customized for' the DAP layer is based on the comparison FT-RLM-DAP vs Self-DANA. These two conditions differ in two ways: PT-RLM pre-training always supplies 12 channels with zeroed masked leads, whereas PT-RLS supplies only the selected leads, and FT-RLM-DAP therefore suffers a train/test distribution shift that FT-RLM-ZP does not. The lower scores of FT-RLM-DAP could be explained by this shift rather than by RLS being intrinsically better suited to DAP. Since RLM and RLS induce the same distribution over retained channel subsets (uniform number of retained leads from 1 to 12 and uniform subset of that size), the operative difference is zeroing versus removal. Please add a control that matches the pre-training input format, such as a lead-removal variant of RLM, or explicitly restrict the claim to 'RLS with DAP outperforms RLM with DAP in our setting'.
  3. [§4.1(iii), Table 3] The comparison between Self-DANA and the channel-specific supervised models is not sufficiently specified to be reproducible. Section 3.5(iii) states that a dedicated fully supervised model was trained for each configuration, but the appendix does not report the supervised training procedure: whether the same backbone with the DAP layer was used, what hyperparameters, number of epochs, early stopping criteria, or data sampling strategy were employed. Without these details, the claimed advantage of a channel-adaptive FM over channel-specific supervised models cannot be evaluated or replicated.
minor comments (5)
  1. [§2 vs §3.2] The related-work section says [18] masks 6 out of 12 randomly chosen channels, while Section 3.2 defines RLM as masking a uniformly random number of channels between 0 and 11. Please clarify whether the implementation used in this paper follows [18] exactly or modifies the masking distribution.
  2. [§3.4 vs Appendix A.1] Section 3.4 states that the Georgia dataset consists of 9,458 12-lead ECGs, while Appendix A.1 reports 20,672 total recordings (10,344 train, 5,167 validation, 5,161 test) and says that only the training set was used. The relationship between the 9,458 figure and the 10,344/5,167/5,161 figures should be clarified.
  3. [Table 4] The peak CPU memory for Self-DANA is identical for the 2-lead and 1-lead configurations (18.00 MB). Since the input size differs between these configurations, please explain why the memory plateau occurs, for example by reporting the memory breakdown or the dominant fixed cost.
  4. [§3.1] The text says the DAP layer 'reduces any input dimension (C, T) to (1, 156)'. The architecture table shows the reduction is applied after the convolutional feature encoder, not directly to the raw input, and the output is [B, 256, 1, 156] before flattening. Please make this description consistent with the architecture.
  5. [§4.1(ii)] The comparison with the literature values from [18] and [34] is acknowledged to be non-fair because of different test data and in-domain pre-training. This caveat is appropriate, but the term 'state-of-the-art' in the abstract should be softened or justified with a directly comparable baseline rather than the literature values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported results are empirical comparisons on held-out Georgia test data, and the architectural choices are adopted or proposed rather than derived from the target metric.

full rationale

The paper contains no derivation chain in which a claimed prediction reduces to a fitted input or to a prior result by the same authors. The DAP layer is presented as an adopted architectural component from prior work (DANA), and Random Lead Selection is introduced as a new stochastic augmentation; neither is defined in terms of the downstream CinC scores. All performance numbers come from fine-tuning on the Georgia training split and evaluating on a held-out Georgia test split, so the central empirical claims are not self-definitional. Hyperparameters such as temperature, learning rate, and augmentation ranges are taken from cited prior work and are not tuned on the test set. The paper explicitly notes that the 12-lead equality between zero-padding and DAP is 'equivalent by design', which is a transparency statement about a control condition, not a circular prediction. It also explicitly cautions that literature comparisons are not fair because prior models were pre-trained on Georgia, showing awareness of evaluation leakage rather than exploiting it. The claim that RLS is 'customized' for DAP is an empirical conclusion from ablations, not a logical identity. Any concerns about the RLM-versus-RLS comparison being confounded are experimental-design limitations, not circularity, and they are outside the scope of this pass. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed empirical patterns masquerading as derivations. The paper is self-contained against external benchmarks, so the correct circularity score is 0.

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

The method introduces no new physical entities or fitted parameters; it relies on standard SSL training and two architectural/augmentation choices. The main unproven assumptions are domain-level: transferability of pre-trained representations and preservation of diagnostic information by the DAP average pooling.

free parameters (4)
  • SimCLR temperature tau = 0.5
    Set to the value used in [18], not tuned on the downstream test set.
  • Pre-training learning rate and decay = 5e-5, gamma 0.97
    Adopted from the reference architecture; not fitted to the Georgia dataset.
  • Fine-tuning learning rate and decay = 1e-5, gamma 0.97
    Selected by the authors; no search reported, and it does not condition on test labels.
  • RLS number of retained leads = uniform in [1,12]
    Chosen to cover all reduced-lead scenarios; the exact distribution is a design choice that affects pre-training but is not fitted to any target metric.
assumptions (4)
  • domain assumption A model pre-trained on 855K windows from seven datasets transfers to the out-of-domain Georgia dataset.
    The paper relies on the transferability of self-supervised representations across ECG datasets with different distributions, which is empirically common but not guaranteed.
  • domain assumption Average pooling across the channel dimension in the DAP layer preserves task-relevant diagnostic information.
    The architecture is forced to summarize all leads into a single 156-length feature vector per time step; the paper assumes lead-specific details are not essential for the 23-class classification task.
  • domain assumption The random distribution of lead subsets used in RLS is sufficient to make the model robust to the fixed reduced-lead configurations evaluated at test time.
    The downstream configurations are specific (e.g., 6-lead set I, II, III, aVF, aVL, aVR), while RLS samples all subsets uniformly; the paper assumes this coverage is adequate.
  • domain assumption Self-supervised contrastive pre-training with SimCLR and the selected augmentations learns useful ECG representations.
    This is supported by prior work, but the paper does not independently prove it; it is a background assumption of the SSL framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-DANA: A Resource-Efficient Channel-Adaptive Self-Supervised Approach for ECG Foundation Models." pith.science (2026). https://pith.science/paper/ARPSRJXG

@misc{pith2026250714151,
  author       = {Pith},
  title        = {Pith review of: Self-DANA: A Resource-Efficient Channel-Adaptive Self-Supervised Approach for ECG Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ARPSRJXG}},
  note         = {Machine review of arXiv:2507.14151}
}
read the original abstract

Foundation Models (FMs) are large-scale machine learning models trained on extensive, diverse datasets that can be adapted to a wide range of downstream tasks with minimal fine-tuning. In the last two years, interest in FMs has also grown for applications in the cardiological field to analyze the electrocardiogram (ECG) signals. One of the key properties of FMs is their transferability to a wide range of downstream scenarios. With the spread of wearable and portable devices, keen interest in learning from reduced-channel configurations has arisen. However, the adaptation of ECG FMs to downstream scenarios with fewer available channels still has to be properly investigated. In this work, we propose Self-DANA, a novel, easy-to-integrate solution that makes self-supervised architectures adaptable to a reduced number of input channels, ensuring resource efficiency and high performance. We also introduce Random Lead Selection, a novel augmentation technique to pre-train models in a more robust and channel-agnostic way. Our experimental results on five reduced-channel configurations demonstrate that Self-DANA significantly enhances resource efficiency while reaching state-of-the-art performance. It requires up to 69.3% less peak CPU memory, 34.4% less peak GPU memory, about 17% less average epoch CPU time, and about 24% less average epoch GPU time.

Figures

Figures reproduced from arXiv: 2507.14151 by the authors.

Figure 1
Figure 1. Overview of the experimental design, including the pre-training and fine-tuning procedures. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LAEF: A Lead-Agnostic ECG Foundation Model Towards Point-of-Care Diagnostics

    cs.LG 2026-08 conditional novelty 7.0 of 10

    A 7M-parameter graph ECG foundation model pre-trained with random lead dropout matches 12-lead models on full input and beats zero-padded baselines on most datasets with 1-2 leads.

Reference graph

Works this paper leans on

42 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stef...

  2. [2]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova Google, and A I Language. Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North, pages 4171–4186, 2019. doi: 10.18653/V1/N19-1423. URL https://aclanthology.org/N19-1423/

  3. [3]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwi...

  4. [4]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. Pro- ceedings of Machine Learning Research , 139:8748–8763, 2 2021. ISSN 26403498. URL https://arx...

  5. [5]

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

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neu- ral Information Processing Systems , 2020-December, 6 2020. ISSN 10495258. URL https://arxiv.org/pdf/2006.11477

  6. [6]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units

    Wei Ning Hsu, Benjamin Bolte, Yao Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio Speech and Language Processing, 29:3451–3460, 6 2021. ISSN 23299304. doi: 10.1109/TASLP.2021.3122291. URL https://arx...

  7. [7]

    Sleepfm: Multi-modal representation learning for sleep across brain activity, ecg and respiratory signals

    Rahul Thapa, Bryan He, Magnus Ruud Kjær, Hyatt Moore, Gauri Ganjoo, Emmanuel Mignot, and James Zou. Sleepfm: Multi-modal representation learning for sleep across brain activity, ecg and respiratory signals. Proceedings of Machine Learning Research, 235:48019–48037, 5

  8. [8]

    Nadkarni

    Akhil Vaid, Joy Jiang, Ashwin Sawant, Stamatios Lerakis, Edgar Argulian, Yuri Ahuja, Joshua Lampert, Alexander Charney, Hayit Greenspan, Jagat Narula, Benjamin Glicks- berg, and Girish N. Nadkarni. A foundational vision transformer improves diagnos- tic performance for electrocardiograms. npj Digital Medicine , 6:1–8, 12 2023. ISSN 23986352. doi: 10.1038/...

Show all 42 references
  1. [9]

    Applications of self-supervised learning to biomedical signals: A survey

    Federico Del Pup and Manfredo Atzori. Applications of self-supervised learning to biomedical signals: A survey. IEEE Access, 11:144180–144203, 2023. ISSN 21693536. doi: 10.1109/ ACCESS.2023.3344531

  2. [10]

    Foundation models in electrocardiogram: A review, 2024

    Yu Han, Xiaofeng Liu, Xiang Zhang, and Cheng Ding. Foundation models in electrocardiogram: A review, 2024. URL https://arxiv.org/abs/2410.19877

  3. [11]

    Ecg-fm: An open electrocardiogram foundation model, 2024

    Kaden McKeen, Laura Oliva, Sameer Masood, Augustin Toma, Barry Rubin, and Bo Wang. Ecg-fm: An open electrocardiogram foundation model, 2024. URL https://arxiv.org/ abs/2408.05178

  4. [12]

    Miller, Saba Emrani, Udhyakumar Nal- lasamy, and Ian Shapiro

    Salar Abbaspourazad, Oussama Elachqar, Andrew C. Miller, Saba Emrani, Udhyakumar Nal- lasamy, and Ian Shapiro. Large-scale training of foundation models for wearable biosignals. 12th International Conference on Learning Representations, ICLR 2024 , 12 2023. URL https://arxiv.o...

  5. [13]

    Ecg semantic integrator (esi): A foundation ecg model pretrained with llm-enhanced cardiological text, 2024

    Han Yu, Peikun Guo, and Akane Sano. Ecg semantic integrator (esi): A foundation ecg model pretrained with llm-enhanced cardiological text, 2024. URL https://arxiv.org/abs/2405. 19366

  6. [14]

    Foundation mod- els for cardiovascular disease detection via biosignals from digital stethoscopes

    George Mathew, Daniel Barbosa, John Prince, and Subramaniam Venkatraman. Foundation mod- els for cardiovascular disease detection via biosignals from digital stethoscopes. npj Cardiovas- cular Health 2024 1:1, 1:1–13, 10 2024. ISSN 2948-2836. doi: 10.1038/s44325-024-00027-5. U...

  7. [15]

    Remote and wearable ecg devices with diagnostic abilities in adults: a state-of-the-science scoping review

    Zeineb Bouzid, Salah S Al-Zaiti, Raymond Bond, and Ervin Sejdi´c. Remote and wearable ecg devices with diagnostic abilities in adults: a state-of-the-science scoping review. Heart Rhythm, 19(7):1192–1201, 2022

  8. [16]

    3kg: Contrastive learning of 12-lead electrocardiograms using physiologically-inspired augmen- tations

    Bryan Gopal, Ryan Han, Gautham Raghupathi, Andrew Ng, Geoff Tison, and Pranav Rajpurkar. 3kg: Contrastive learning of 12-lead electrocardiograms using physiologically-inspired augmen- tations. Proceedings of Machine Learning Research, 158:156–167, 4 2021. ISSN 26403498. URL ht...

  9. [17]

    Dense lead contrast for self-supervised representation learning of multilead electrocardiograms

    Wenhan Liu, Zhoutong Li, Huaicheng Zhang, Sheng Chang, Hao Wang, Jin He, and Qi- jun Huang. Dense lead contrast for self-supervised representation learning of multilead electrocardiograms. Information Sciences, 634:189–205, 7 2023. ISSN 0020-0255. doi: 10.1016/J.INS.2023.03.09...

  10. [18]

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

    Jungwoo Oh, Hyunseung Chung, Dong-Gyun Hong, and Edward Choi. Lead-agnostic self- supervised learning for local and global representations of electrocardiogram. Proceed- ings of Machine Learning Research, 174:2022, 2022. URL https://github.com/Jwoo5/ fairseq-signals

  11. [19]

    Clegg, Andrea Cavallaro, and Hamed Haddadi

    Mohammad Malekzadeh, Richard G. Clegg, Andrea Cavallaro, and Hamed Haddadi. Dana: Dimension-adaptive neural architecture for multivariate sensor data. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 5:120, 8 2020. doi: 10.1145/ 3478074. URL...

  12. [20]

    An electrocardiogram foundation model built on over 10 million recordings with external evaluation across multiple domains, 2025

    Jun Li, Aaron Aguirre, Junior Moura, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Clifford, Brandon Westover, and Shenda Hong. An electrocardiogram foundation model built on over 10 million recordings with external evaluation across multiple domains, 2025. URL https: //arxiv.org/ab...

  13. [21]

    Biot: Biosignal transformer for cross-data learning in the wild

    Chaoqi Yang, M Brandon Westover, and Jimeng Sun. Biot: Biosignal transformer for cross-data learning in the wild. Advances in Neural Information Processing Systems, 36:78240–78260, 12 2023

  14. [22]

    Will two do? varying dimensions in electrocardiography: The physionet/computing in cardiology challenge 2021

    Matthew A Reyna, Nadi Sadr, Erick A Perez Alday, Annie Gu, Amit J Shah, Chad Robichaux, Ali Bahrami Rad, Andoni Elola, Salman Seyedi, Sardar Ansari, Hamid Ghanbari, Qiao Li, Ashish Sharma, and Gari D Clifford. Will two do? varying dimensions in electrocardiography: The physion...

  15. [23]

    Issues in the automated classification of multilead ecgs using heterogeneous labels and populations

    Matthew A Reyna, Nadi Sadr, Erick A Perez Alday, Annie Gu, Amit J Shah, Chad Robichaux, Ali Bahrami Rad, Andoni Elola, Salman Seyedi, Sardar Ansari, Hamid Ghanbari, Qiao Li, Ashish Sharma, and Gari D Clifford. Issues in the automated classification of multilead ecgs using hete...

  16. [24]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 37th International Conference on Machine Learning, ICML 2020, PartF168147-3:1575–1585, 2 2020. URL https://arxiv.org/pdf/ 2002.05709

  17. [25]

    Ribeiro, Manoel Horta Ribeiro, Gabriela M.M

    Antônio H. Ribeiro, Manoel Horta Ribeiro, Gabriela M.M. Paixão, Derick M. Oliveira, Paulo R. Gomes, Jéssica A. Canazart, Milton P.S. Ferreira, Carl R. Andersson, Peter W. Macfarlane, Meira Wagner, Thomas B. Schön, and Antonio Luiz P. Ribeiro. Automatic diagnosis of the 12-lead...

  18. [26]

    Ribeiro, Gabriela M.M

    Antônio H. Ribeiro, Gabriela M.M. Paixao, Emilly M. Lima, Manoel Horta Ribeiro, Marcelo M. Pinto Filho, Paulo R. Gomes, Derick M. Oliveira, Wagner Meira Jr, Thömas B Schon, and Antonio Luiz P. Ribeiro. Code-15 URL https://zenodo.org/records/4916206. v1.0.0; Accessed: 2024-01-1...

  19. [27]

    Practical intelligent diagnostic algorithm for wearable 12-lead ecg via self- supervised learning on large-scale dataset

    Jiewei Lai, Huixin Tan, Jinliang Wang, Lei Ji, Jun Guo, Baoshi Han, Yajun Shi, Qianjin Feng, and Wei Yang. Practical intelligent diagnostic algorithm for wearable 12-lead ecg via self- supervised learning on large-scale dataset. Nature Communications 2023 14:1 , 14:1–13, 6

  20. [28]

    URL https://www.scidb.cn/en/ detail?dataSetId=58c4a92d5a01414390a78160d335380d

    Offline test set of ecg multi-label classfication, 2023. URL https://www.scidb.cn/en/ detail?dataSetId=58c4a92d5a01414390a78160d335380d. v1; Accessed: 2024-05-02; MIT license

  21. [29]

    A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(...

  22. [30]

    Will two do? varying dimensions in electrocardiography: The physionet/computing in cardiology challenge 2021

    Matthew A Reyna, Nadi Sadr, Annie Gu, Erick A Perez Alday, Chengyu Liu, Salman Seyedi, Amit J Shah, and Gari D Clifford. Will two do? varying dimensions in electrocardiography: The physionet/computing in cardiology challenge 2021. PhysioNet, 2022. URL https: //physionet.org/co...

  23. [31]

    Optimal multi-stage arrhythmia classification approach

    Jianwei Zheng, Huimin Chu, Daniele Struppa, Jianming Zhang, Magdi Yacoub, Hesham El- Askary, Anthony Chang, Louis Ehwerhemuepha, Islam Abudayyeh, Alexander Barrett, Guohua Fu, Hai Yao, Dongbo Li, Hangyuan Guo, and Cyril Rakovski. Optimal multi-stage arrhythmia classification a...

  24. [32]

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

    Jianwei Zheng, Jianming Zhang, Sidy Danioko, Hai Yao, Hangyuan Guo, and Cyril Rakovski. A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients. Scientific Data, 7, 02 2020. doi: 10.1038/s41597-020-0386-x. 24

  25. [33]

    An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detection

    Feifei Liu, Chengyu Liu, Lina Zhao, Xiangyu Zhang, Xiaoling Wu, Xiaoyan Xu, Yulin Liu, Caiyun Ma, Shoushui Wei, Zhiqiang He, Jianqing Li, Eddie Ng, and Yin Kwee. An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detect...

  26. [34]

    Classification of ecg using ensemble of residual cnns with attention mechanism

    Petr Nejedly, Adam Ivora, Radovan Smisek, Ivo Viscor, Zuzana Koscova, Pavel Jurak, and Filip Plesinger. Classification of ecg using ensemble of residual cnns with attention mechanism. Computing in Cardiology, 48, 2021. doi: 10.22489/CinC.2021.014

  27. [35]

    Analysis of augmentations for contrastive ecg representation learning

    Sahar Soltanieh, Ali Etemad, and Javad Hashemi. Analysis of augmentations for contrastive ecg representation learning. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1–10, 2022. doi: 10.1109/IJCNN55064.2022.9892600

  28. [36]

    Self-supervised representation learning from 12-lead ecg data

    Temesgen Mehari and Nils Strodthoff. Self-supervised representation learning from 12-lead ecg data. Computers in Biology and Medicine, 141:105114, 2 2022. ISSN 0010-4825. doi: 10. 1016/J.COMPBIOMED.2021.105114. URL https://www.sciencedirect.com/science/ article/pii/S0010482521009082

  29. [37]

    Federico Del Pup, Andrea Zanola, Louis Fabrice Tshimanga, Paolo Emilio Mazzon, and Manfredo Atzori. selfeeg. GitHub repository, 2024. URLhttps://github.com/MedMaxLab/ selfEEG. v0.2.0; Accessed: 2024-10-09; MIT license

  30. [38]

    Selfeeg: A python library for self-supervised learning in electroencephalog- raphy

    Federico Del Pup, Andrea Zanola, Louis Fabrice Tshimanga, Paolo Emilio Mazzon, and Manfredo Atzori. Selfeeg: A python library for self-supervised learning in electroencephalog- raphy. Journal of Open Source Software, 9(95):6224, 2024. doi: 10.21105/joss.06224. URL https://doi....

  31. [39]

    GitHub repository, 2023

    fairseq-signals. GitHub repository, 2023. URL https://github.com/Jwoo5/ fairseq-signals. Accessed: 2024-10-09; MIT license. 25

  32. [2022]

    URL https://arxiv.org/abs/2108.07258

  33. [2023]

    doi: 10.1038/s41467-023-39472-8

    ISSN 2041-1723. doi: 10.1038/s41467-023-39472-8. URL https://www.nature. com/articles/s41467-023-39472-8

  34. [2024]

    URL https://arxiv.org/pdf/2405.17766

    ISSN 26403498. URL https://arxiv.org/pdf/2405.17766. 22

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.