Pith. sign in

REVIEW 3 major objections 5 minor 29 references

LSTrans classifies multi-lead ECGs with competitive diagnostic sensitivity while cutting peak memory and training latency far below heavier models.

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 · grok-4.5

2026-07-14 09:16 UTC pith:L6IIWXN4

load-bearing objection Solid engineering paper: a coherent 1D interleaved student + dual-rank LoRA + hetero KD that actually delivers competitive multi-label ECG numbers at low training memory; the only real soft spot is the unmeasured edge-inference claim. the 3 major comments →

arxiv 2607.10784 v1 pith:L6IIWXN4 submitted 2026-07-12 cs.LG cs.AI

LSTrans: Efficient Knowledge Transfer for Lightweight and Automated ECG Classification

Yi Zhao , Jiajun Gao , Chenyang Xu , Yuxi Zhou , Hao Wang This is my paper
classification cs.LG cs.AI
keywords ECGCVDsDeep LearningLightweight ModelsEfficient DiagnosisKnowledge DistillationLoRATransformer
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.

The paper sets out to show that automated multi-label ECG diagnosis can be made practical for resource-limited wearable devices without sacrificing the ability to catch subtle morphological and rhythmic abnormalities. It introduces LSTrans, a hybrid student model whose 1D interleaved convolutional backbone extracts both local wave shapes and longer rhythm context, then hands those features to a Transformer that is kept small by Low-Rank Adaptation. Larger teacher networks transfer their diagnostic knowledge through homogeneous or heterogeneous distillation so the compact student retains high recall. On three standard 12-lead benchmarks the resulting system reaches average AUC near 0.95 and strong Fβ=2 scores while using roughly 0.86 GB peak memory and 0.084 s per training step. Anyone interested in continuous, out-of-hospital cardiac monitoring would care because the work claims a usable accuracy–efficiency trade-off that earlier heavy networks and even some specialized lightweight baselines do not achieve.

Core claim

LSTrans (especially the heterogeneous-distillation plus LoRA variant) reaches competitive diagnostic sensitivity—average AUC 0.949 and Fβ=2 of 0.716 across G12EC, PTB-XL and Chapman-Shaoxing—while substantially lowering peak GPU memory and per-iteration training latency relative to foundation models and other efficient baselines, by combining a specialized interleaved 1D convolutional backbone, a LoRA-adapted Transformer, and structured knowledge distillation.

What carries the argument

The 1D-LSNet interleaved backbone: even-indexed layers use reparameterized multi-scale depthwise convolutions plus Squeeze-and-Excitation for local morphology, odd-indexed layers use Large-Kernel Perception / Small-Kernel Aggregation for macro-rhythms (with late multi-head attention), cascaded into a LoRA-compressed Transformer and trained under homogeneous or heterogeneous knowledge distillation.

Load-bearing premise

The paper treats peak GPU memory and training-step latency measured on a high-end desktop GPU as sufficient evidence that the model will be deployable on actual low-power wearable hardware—an assumption it itself flags as untested.

What would settle it

Deploy the final student model on a representative wearable microcontroller or edge SoC, measure real-time inference latency, power draw and multi-label accuracy on continuous multi-lead ECG streams; if energy or latency exceeds clinical wearable budgets while Fβ=2 falls below the reported levels, the efficiency claim fails for the intended setting.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Multi-label clinical ECG classifiers become feasible on memory-constrained wearable nodes without full foundation-model footprints.
  • Downstream fine-tuning of ECG networks can stay under 1 GB peak memory while preserving high recall-oriented metrics.
  • Heterogeneous distillation from residual teachers can restore diagnostic sensitivity lost during aggressive compression.
  • Dual-rank LoRA (minimal rank on convolutions, higher rank on attention) can act simultaneously as a parameter saver and a regularizer against noise and catastrophic forgetting.

Where Pith is reading between the lines

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

  • The same dual-scale interleaved design is likely to transfer to other non-stationary 1D biosignals such as PPG or EEG where both local morphology and long-range rhythms matter.
  • If the dynamic weight generation inside Large-Kernel Perception proves costly on microcontrollers, static reparameterization of that path could close the remaining edge-deployment gap the authors note.
  • Clinical validation on actual wearable ECG patches—already listed by the authors as future work—would be the decisive real-world gate for adoption beyond benchmark datasets.

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

3 major / 5 minor

Summary. The paper proposes LSTrans, a lightweight hybrid ECG classifier that combines a specialized 1D interleaved convolutional backbone (1D-LSNet with alternating RepVGGDW+SE and LKP+SKA blocks) with a Transformer encoder, LoRA adapters (dual ranks rc=4 / rt=16), and both homogeneous and heterogeneous knowledge distillation. The central claim is that the Hetero+LoRA student achieves competitive multi-label diagnostic sensitivity (average AUC ≈ 0.949, Fβ=2 ≈ 0.716 across G12EC, PTB-XL and Chapman-Shaoxing) while substantially reducing peak training memory (0.856 GB) and per-iteration latency (0.084 s) relative to heavier baselines and a foundation model (Table I). The work is supported by large-scale pretraining on CODE-15%, patient-disjoint 10-fold CV, component ablations (Table II, Figs. 3–5), and released code.

Significance. If the reported efficiency–accuracy trade-off holds, the work is a useful engineering contribution for multi-label 12-lead ECG on resource-constrained platforms. Strengths include the explicit dual-scale interleaved design, dual-rank LoRA justified by intrinsic-dimensionality arguments, systematic homo/hetero distillation comparison, patient-disjoint evaluation on three public benchmarks, and open-source code. The training-memory and latency reductions versus H-Tuning, AdaGrad-Fusion and ECG-Founder are practically relevant for fine-tuning pipelines even if edge inference remains unmeasured. The contribution is incremental rather than foundational, but the combination of architecture, PEFT and distillation is carefully executed and reproducible.

major comments (3)
  1. [Table I / §III-C] Table I and §III-C: several key baselines (AdaGrad-Fusion, ECG-Founder) are reported without standard deviations or fold-level statistics, while the proposed Hetero+LoRA variants include them. Given that the central claim rests on matching or exceeding these baselines at lower cost, the comparison is incomplete; either re-run those baselines under the same 10-fold protocol or clearly mark them as single-run numbers and temper the superiority language.
  2. [Abstract / §IV] Abstract, Introduction and Limitations: the framing repeatedly invokes “resource-constrained wearable devices” and “IoHT nodes,” yet all efficiency numbers are peak GPU memory and training iteration time on an RTX 4090. The limitations section itself acknowledges that inference latency on non-GPU edge hardware and the control-flow cost of dynamic LKP/SKA kernels are unmeasured. Either supply at least one edge-device (CPU/MCU/NPU) inference measurement or re-scope the claim to “training-time / fine-tuning efficiency,” which is what the experiments actually support.
  3. [§III-A] §III-A: discarding all CVD classes with fewer than 200 samples is a free modeling choice that can inflate multi-label metrics by removing the hardest rare labels. The paper should report the number of retained vs. discarded classes per dataset and, ideally, a sensitivity check (or at least macro metrics on the full label set) so readers can judge whether the reported Fβ=2 gains partly arise from this filter.
minor comments (5)
  1. [Eq. (1) / §II-A] Eq. (1) and surrounding text: the temperature scaling is written as T^{2} · ℓ(σ(f_s/T), σ(f_t/T)), which is standard, but α and T values used in the final runs are never stated; please list them for reproducibility.
  2. [Fig. 1] Fig. 1 caption and panel (c) still show “BN-Conv 2D” and “K=1×9” while the method is 1D; this is confusing and should be corrected to 1D notation.
  3. [Table II] Table II memory numbers for the ablated variants (0.537–0.827 GB) are lower than the full Hetero+LoRA entry in Table I (0.856 GB); a short note explaining the difference (e.g., presence/absence of the Transformer head) would help.
  4. [§III-B] Several secondary metrics (HL, ranking loss, coverage, MAP, Gβ=2) are mentioned in §III-B and partially discussed in text but never tabulated; either add a supplementary table or drop the unused names.
  5. [Throughout] Typographical issues: “joy yuxi@pku.edu.cn” (space), “Gr o u nd Truth” in Fig. 1, and inconsistent capitalization of “LSTrans hetero” vs. “LSTrans(Hetero+LoRA)” throughout.

Circularity Check

0 steps flagged

No circularity: purely empirical architecture + distillation results measured on held-out folds against external baselines.

full rationale

LSTrans is an engineering paper whose central claims (competitive macro-AUC / Fβ=2 at reduced peak training memory and iteration latency) rest on patient-disjoint 10-fold cross-validation numbers, ablations of the interleaved 1D-LSNet blocks, dual-rank LoRA, and homogeneous/heterogeneous KD (Tables I–II, Eq. 1). Soft-label temperature T and LoRA ranks rc=4 / rt=16 are free hyperparameters, not quantities fitted to the target metrics and then re-reported as predictions. Self-comparisons to the authors’ own ResNet/LSNet teachers and to one overlapping-author baseline (AdaGrad-Fusion) are ordinary experimental controls; none of them supply a uniqueness theorem, ansatz, or definitional identity that forces the reported gains. The derivation chain therefore contains no self-definitional step, no fitted-input-as-prediction, and no load-bearing self-citation. The paper is self-contained against external public ECG benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

An empirical systems paper whose claims rest on architectural design choices, a handful of free hyper-parameters, and standard domain assumptions about ECG morphology rather than on unproved mathematical lemmas or new physical entities.

free parameters (4)
  • LoRA ranks rc=4 (conv), rt=16 (Transformer)
    Chosen by hand from an intrinsic-dimension argument; directly control trainable parameter count and the reported memory/accuracy trade-off.
  • distillation weight α and temperature T
    Balance hard-label BCE against soft-label KD loss (Eq. 1); values affect the sensitivity gains claimed for the student.
  • stage depths d ∈ [1,2,8,10] and channel progression
    Define the capacity of the 1D-LSNet backbone; selected without exhaustive search reported.
  • learning rates 0.001 (teacher) / 0.002 (student), batch 64, patience 30
    Optimization hyper-parameters that determine the final numbers in Table I.
axioms (4)
  • domain assumption ECG signals contain complementary macroscopic rhythmic trends and microscopic morphological variations that can be captured by alternating large-kernel and multi-scale local filters.
    Stated in §II-B and used to justify the even/odd interleaved block design.
  • domain assumption Soft probability distributions from a high-capacity teacher transfer ‘dark knowledge’ that restores diagnostic sensitivity under parameter constraints.
    Standard KD premise invoked in §II-E and Eq. 1.
  • ad hoc to paper Convolutional layers have low intrinsic dimensionality while Transformer attention maps have high intrinsic dimensionality, justifying dual-rank LoRA.
    Cited from Ansuini et al. but applied without new measurement on ECG features (§II-D).
  • ad hoc to paper Classes with fewer than 200 samples can be discarded without biasing multi-label metrics.
    Explicit filtering rule in §III-A.
invented entities (2)
  • 1D-LSNet interleaved backbone (RepVGGDW+SE alternating with LKP+SKA) no independent evidence
    purpose: Extract dual-scale ECG features while remaining parameter-light.
    Architectural construct introduced in §II-B; no independent physiological validation outside the paper’s own ablations.
  • dual-rank LoRA allocation (rc=4 / rt=16) specialized for ECG no independent evidence
    purpose: Compress trainable space while regularizing local vs. global features.
    Paper-specific schedule; evidence is only the reported memory/accuracy numbers.

pith-pipeline@v1.1.0-grok45 · 19087 in / 3042 out tokens · 57830 ms · 2026-07-14T09:16:54.955096+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of LSTrans: Efficient Knowledge Transfer for Lightweight and Automated ECG Classification." pith.science (2026). https://pith.science/paper/L6IIWXN4

@misc{pith2026260710784,
  author       = {Pith},
  title        = {Pith review of: LSTrans: Efficient Knowledge Transfer for Lightweight and Automated ECG Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L6IIWXN4}},
  note         = {Machine review of arXiv:2607.10784}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deploying deep learning models for automated electrocardiogram classification on resource-constrained wearable devices remains challenging due to high computational costs. To address this, we propose LSTrans, a lightweight hybrid model designed for efficient and sensitive ECG analysis. LSTrans introduces a specialized 1D convolutional backbone with an interleaved layer architecture to capture both macroscopic rhythmic trends and microscopic morphological variations. This backbone is cascaded with a Transformer encoder to model long-range temporal dependencies, incorporating Low-Rank Adaptation across critical layers to compress the model and reduce the trainable parameter space. We further employ homogeneous and heterogeneous knowledge distillation to transfer diagnostic expertise from high-capacity teacher models to the student. Experimental results on multiple benchmark datasets demonstrate that LSTrans achieves a competitive balance between diagnostic sensitivity and resource efficiency, substantially reducing peak memory footprints and training latency during downstream adaptation. The source code is available for review at https://github.com/zyee00128/LSTrans4BIBM.

Figures

Figures reproduced from arXiv: 2607.10784 by Chenyang Xu, Hao Wang, Jiajun Gao, Yi Zhao, Yuxi Zhou.

Figure 1
Figure 1. Figure 1: Overview of the knowledge distillation pipeline. (a) Homogeneous [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The diagram of the LSTrans heterogeneous framework contains the preprocessing and joint distillation training pipeline, hyperparameter sensitivity [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: SE Gate Activation Weights 4 3 2 1 0 1 2 Amplitude SKA Feature Enhancement Analysis Raw ECG Signal 2 0 2 4 6 Activation Before SKA 0 50 100 150 200 250 Time (Samples) 5 0 5 10 Activation After SKA [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SKA Feature Enhancement Analysis As an illustrative example, Left Bundle Branch Block (LBBB) clinically manifests as deep S-waves in septal leads V1 and V2, alongside broad R-waves in lateral leads V5 and V6. The model assigns prominent attention weights 0.812 to lead V5 and 0.761 to lead V2 under the LBBB category. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LKP Kernel Weights TABLE II ABLATION STUDY ON LSNET IN G12EC DATASET. Model AUC Fβ=2 Mem Time w/o SKA 0.905 ± 0.000 0.653 ± 0.000 0.569 0.079 w/o LKP 0.905 ± 0.001 0.655 ± 0.005 0.537 0.076 w/o LSConv 0.903 ± 0.002 0.651 ± 0.010 0.680 0.087 w/o RepVGG 0.906 ± 0.002 0.650 ± 0.002 0.656 0.083 w/o se 0.902 ± 0.002 0.649 ± 0.007 0.655 0.083 w/o even-indexed layers 0.902 ± 0.001 0.643 ± 0.001 0.827 0.100 LSTran… 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

29 extracted references · 6 linked inside Pith

  1. [1]

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

    A. Y . Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourn, M. P. Turakhia, and A. Y . Ng, “Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network,”Nature medicine, vol. 25, no. 1, pp. 65–69, 2019

    2019

  2. [2]

    Automatic diagnosis of the 12-lead ecg using a deep neural network,

    A. H. Ribeiro, M. H. Ribeiro, G. M. Paix ˜ao, D. M. Oliveira, P. R. Gomes, M. R. Canesin, B. D. Nearing, and A. L. Goldberger, “Automatic diagnosis of the 12-lead ecg using a deep neural network,”Nature communications, vol. 11, no. 1, pp. 1–9, 2020

    2020

  3. [3]

    A systematic review on the use of consumer-based ecg wearables on cardiac health monitoring,

    R. Wang, S. C. M. Veera, O. Asan, and T. Liao, “A systematic review on the use of consumer-based ecg wearables on cardiac health monitoring,” IEEE Journal of Biomedical and Health Informatics, 2024

    2024

  4. [4]

    Wearable devices for remote monitoring of chronic diseases: Systematic review,

    M. D. Tegegne, S. R. Niakan Kalhori, P. Haas, V . M. Sobotta, J. War- necke, and T. M. Deserno, “Wearable devices for remote monitoring of chronic diseases: Systematic review,”JMIR mHealth and uHealth, vol. 14, p. e74071, 2026

    2026

  5. [5]

    Explainable deep learning based techniques for ecg-based heart disease classification: A systematic literature review and future direction,

    G. Manimaran, A. Peimankar, S. Puthusserypady, M. Momeni, R. A. I. Asyari, M. S. Jahan, J. Moll, U. K. Wiil, and A. Ebrahimi, “Explainable deep learning based techniques for ecg-based heart disease classification: A systematic literature review and future direction,”Computers in Biology and Medicine, vol. 199, p. 111324, 2025

    2025

  6. [6]

    Advanced deep learning framework for ecg arrhythmia classification using 1d-cnn with attention mechanism,

    M. Guhdar, A. O. Mohammed, and R. J. Mstafa, “Advanced deep learning framework for ecg arrhythmia classification using 1d-cnn with attention mechanism,”Knowledge-Based Systems, vol. 315, p. 113301, 2025

    2025

  7. [7]

    Foundation models in electrocardiogram: A review,

    Y . Han, X. Liu, X. Zhang, and C. Ding, “Foundation models in electrocardiogram: A review,”arXiv preprint arXiv:2410.19877, 2024

    arXiv 2024

  8. [8]

    High-accuracy ecg image interpretation using parameter-efficient lora fine-tuning with multimodal llama 3.2,

    M. Nandakishor and M. Anjali, “High-accuracy ecg image interpretation using parameter-efficient lora fine-tuning with multimodal llama 3.2,” arXiv preprint arXiv:2501.18670, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    H-tuning: Toward low-cost and efficient ecg-based cardiovascular disease detection with pre-trained models,

    R. Zhou, Y . Zhang, and Y . Dong, “H-tuning: Toward low-cost and efficient ecg-based cardiovascular disease detection with pre-trained models,” inForty-second International Conference on Machine Learn- ing, 2025

    2025

  10. [10]

    Deep learning for ecg classification: A comparative study of 1d and 2d representations and multimodal fusion approaches,

    H. Narotamo, M. Dias, R. Santos, A. V . Carreiro, H. Gamboa, and M. Silveira, “Deep learning for ecg classification: A comparative study of 1d and 2d representations and multimodal fusion approaches,” Biomedical Signal Processing and Control, vol. 93, p. 106141, 2024

    2024

  11. [11]

    Deep learning for ecg analysis: Benchmarks and insights from ptb-xl,

    N. Strodthoff, P. Wagner, T. Schaeffter, and W. Samek, “Deep learning for ecg analysis: Benchmarks and insights from ptb-xl,”IEEE journal of biomedical and health informatics, vol. 25, no. 5, pp. 1519–1528, 2020

    2020

  12. [12]

    Ecgtransform: Empowering adaptive ecg arrhythmia classification framework with bidirectional transformer,

    H. El-Ghaish and E. Eldele, “Ecgtransform: Empowering adaptive ecg arrhythmia classification framework with bidirectional transformer,” Biomedical Signal Processing and Control, vol. 89, p. 105714, 2024

    2024

  13. [13]

    Mta- net: A multi-scale temporal-attentive network with semantic-structural fusion for multi-label ecg classification,

    S. Pang, Z. Kan, Z. Zhao, W. Wu, H. Ding, and S. Qiao, “Mta- net: A multi-scale temporal-attentive network with semantic-structural fusion for multi-label ecg classification,” in2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2025, pp. 1–6

    2025

  14. [14]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,”arXiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  15. [15]

    Crossstateecg-lite: Lightweight adaptive thresholding network for dual-state ecg biometrics,

    D. Zheng, J. Feng, and J. Liu, “Crossstateecg-lite: Lightweight adaptive thresholding network for dual-state ecg biometrics,” in2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2025, pp. 3357–3362

    2025

  16. [16]

    Hmt-kd: Lightweight deep learning model for ecg analysis using hierarchical,

    N. Beigzadeh and A. Fathi, “Hmt-kd: Lightweight deep learning model for ecg analysis using hierarchical,”Applied Soft Computing, p. 115038, 2026

    2026

  17. [17]

    Ecg-image-kit: a synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization,

    K. K. Shivashankara, Deepanshi, A. Mehri Shervedani, G. D. Clifford, M. A. Reyna, and R. Sameni, “Ecg-image-kit: a synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization,”Physiological measurement, vol. 45, no. 5, p. 055019, 2024

    2024

  18. [18]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, 2021

    Pith/arXiv arXiv 2021

  19. [19]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,”arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  20. [20]

    Diredi: Distillation and reverse distillation for aiot applications,

    C. Sun, Q. Tong, W. Yang, and W. Zhang, “Diredi: Distillation and reverse distillation for aiot applications,”IEEE Open Journal of the Computer Society, 2024

    2024

  21. [21]

    Intrinsic dimension of data representations in deep neural networks,

    A. Ansuini, A. Laio, J. H. Macke, and D. Zoccolan, “Intrinsic dimension of data representations in deep neural networks,”Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  22. [22]

    Df- mlsl: An effective distillation framework for multi-label single-lead ecg classification,

    Y . Qiang, X. Dong, X. Liu, C. Xu, F. Hu, and R. Wang, “Df- mlsl: An effective distillation framework for multi-label single-lead ecg classification,” in2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2025, pp. 4030–4035

    2025

  23. [23]

    Code-15%: A large scale annotated dataset of 12-lead ecgs,

    A. H. Ribeiro, G. M. Paix ˜ao, E. M. Lima, M. H. Ribeiro, M. M. Pinto Filho, P. R. Gomes, D. M. Oliveira, W. Meira Jr, T. B. Schon, and A. L. P. Ribeiro, “Code-15%: A large scale annotated dataset of 12-lead ecgs,”Zenodo, Jun, vol. 9, pp. 10–5281, 2021

    2021

  24. [24]

    Classification of 12-lead ecgs: the physionet/computing in cardiology challenge 2020,

    E. A. P. Alday, A. Gu, A. J. Shah, C. Robichaux, A.-K. I. Wong, C. Liu, F. Liu, A. B. Rad, A. Elola, S. Seyediet al., “Classification of 12-lead ecgs: the physionet/computing in cardiology challenge 2020,” Physiological measurement, vol. 41, no. 12, p. 124003, 2020

    2020

  25. [25]

    Ptb-xl, a large publicly available electro- cardiography dataset,

    P. Wagner, N. Strodthoff, R.-D. Bousseljot, D. Kreiseler, F. I. Lunze, W. Samek, and T. Schaeffter, “Ptb-xl, a large publicly available electro- cardiography dataset,”Scientific data, vol. 7, no. 1, p. 154, 2020

    2020

  26. [26]

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

    J. Zheng, J. Zhang, S. Danioko, H. Yao, H. Guo, and C. Rakovski, “A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients,”Scientific data, vol. 7, no. 1, p. 48, 2020

    2020

  27. [27]

    Adagrad-fusion: Adaptive gradient fusion for memory-efficient ecg foundation model fine-tuning,

    C. Xu, Y . Zhao, Z. Zhang, and H. Wang, “Adagrad-fusion: Adaptive gradient fusion for memory-efficient ecg foundation model fine-tuning,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 8302–8306

    2026

  28. [28]

    Ce-ssl: Computation-efficient semi-supervised learn- ing for ecg-based cardiovascular diseases detection,

    R. Zhou, L. Clifton, Z. Liu, K. W. Chan, D. A. Clifton, Y .-T. Zhang, and Y . Dong, “Ce-ssl: Computation-efficient semi-supervised learn- ing for ecg-based cardiovascular diseases detection,”arXiv preprint arXiv:2406.14377, 2024

    Pith/arXiv arXiv 2024

  29. [29]

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

    J. Li, A. Aguirre, J. Moura, C. Liu, L. Zhong, C. Sun, G. Clifford, B. Westover, and S. Hong, “An electrocardiogram foundation model built on over 10 million recordings with external evaluation across multiple domains,”arXiv preprint arXiv:2410.04133, 2024

    Pith/arXiv arXiv 2024