Pith. sign in

REVIEW 4 major objections 5 minor 44 references

SemiSegECG: A Multi-Dataset Benchmark for Semi-Supervised Semantic Segmentation in ECG Delineation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read SemiSegECG claims to be the first standardized benchmark for semi-supervised semantic segmentation in ECG delineation, and its results show that transformer backbones outperform convolutional backbones when labels are scarce.

desk verdict A genuinely useful benchmark resource whose headline transformer-vs-CNN claim rests on hyperparameters tuned in one setting; the resource deserves serious refereeing, the claim needs to be softened. read the letter →

arxiv 2507.18323 v2 pith:O2A6PFMW submitted 2025-07-24 cs.CV cs.AIcs.LGeess.SP

classification cs.CVcs.AIcs.LGeess.SP
keywords ECGdelineationsemi-supervisedlearningsemanticsegmentationbenchmarkvisiontransformerconvolutionalnetworkaugmentationcross-domaingeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to establish SemiSegECG as the first systematic benchmark for semi-supervised semantic segmentation in ECG delineation, the task of marking P waves, QRS complexes, and T waves in ECG recordings. It unifies four publicly available delineation datasets plus a large unlabeled dataset into one evaluation protocol, with four label-scarcity ratios, two encoder families, and both same-domain and cross-domain settings. The central empirical claim is that under this protocol semi-supervised algorithms improve delineation over a supervised-only baseline when labels are scarce, and that a vision-transformer backbone outperforms a convolutional backbone. This matters because expert ECG annotation is expensive, so a reliable way to use unlabeled clinical data could lower the cost of accurate waveform segmentation.

What carries the argument

The load-bearing object is the benchmark protocol itself: all leads resampled to 250 Hz, cropped or padded to 10 s, bandpass-filtered and z-scored, with each lead treated as an independent training instance. On top of this, SemiSegECG fixes four labeled-subset ratios (1/16, 1/8, 1/4, 1/2), five semi-supervised algorithms (Mean Teacher, FixMatch, Cross Pseudo Supervision, Regional Contrast, Self-Training++), two encoder-decoder architectures (ResNet-18 and ViT-Tiny with a two-layer FCN decoder), and an ECG-specific augmentation set: random resized cropping as the weak augmentation and a four-operation RandAugment of baseline shift, powerline noise, sine-wave noise, amplitude scaling, and white noise as the strong augmentation. Every comparison the paper makes is defined inside this protocol, so the protocol is what carries the argument.

What would settle it

A concrete test: re-run the same five semi-supervised algorithms on QTDB at the 1/16 label ratio with hyperparameters and augmentations tuned on QTDB rather than LUDB; if the ViT-Tiny advantage over ResNet-18 disappears or the semi-supervised gain over Scratch vanishes, the benchmark's headline comparisons are configuration artifacts rather than method differences.

Watch

Extended reading notes

Core claim

On its own terms, the paper discovers that the choice of backbone dramatically changes whether semi-supervised training pays off. With in-domain unlabeled data, ViT-Tiny consistently outperforms ResNet-18 and most algorithms exceed the supervised Scratch baseline, especially at the 1/16 label ratio. In the cross-domain setting, where PTB-XL supplies out-domain unlabeled data and evaluation is on a merged test set plus a mobile-ECG dataset, only ViT-Tiny benefits from semi-supervised training; every ResNet-based semi-supervised method fails to beat Scratch on the merged in-domain test set. The paper also finds that the model with the highest mIoU is not the one with the lowest clinical interval error, so segmentation accuracy and interval accuracy do not track each other.

Load-bearing premise

The paper assumes that a single set of ECG-specific augmentations and semi-supervised hyperparameters, tuned on LUDB at the 1/16 label ratio with a ResNet-18 backbone, transfers without change to all other databases, label ratios, the ViT backbone, and the cross-domain setting.

Editorial extensions

If this is right

  • New semi-supervised ECG delineation methods can be compared directly on a common protocol with four label ratios and two distribution-shift settings.
  • Practitioners with few labeled ECGs can use the benchmark's augmentation recipes, random resized cropping plus noise-based RandAugment, to improve supervised training.
  • Under label scarcity, semi-supervised training, especially Mean Teacher and Self-Training++, is a viable way to exploit unlabeled clinical ECG archives.
  • Vision-transformer encoders appear to extract more benefit from unlabeled data than convolutional encoders in this setting, at least within the configurations tested.
  • Cross-domain unlabeled data helps a ViT but not ResNet, and best mIoU does not imply best clinical interval error, so multi-metric evaluation is needed.

Reading between the lines

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

  • The authors do not test this, but the same protocol could be lifted to other physiological waveforms (EEG, PPG) where labeled data are scarce; the augmentations would need revalidation per signal type.
  • The cross-domain result suggests the interaction between semi-supervised method and encoder inductive bias is where the action is; a systematic sweep over more encoders would test this without collecting new data.
  • The observed mismatch between mIoU and interval MAE implies that for clinical deployment, the objective itself may need to include interval errors; otherwise a high-mIoU model could still misplace clinically relevant onsets and offsets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces SemiSegECG, a benchmark for semi-supervised ECG delineation built from four labeled public datasets (LUDB, QTDB, ISP, Zhejiang), PTB-XL as an unlabeled out-domain source, and a private mobile ECG test set. It evaluates five semi-supervised segmentation algorithms (MT, FixMatch, CPS, ReCo, ST++) plus a supervised Scratch baseline under ResNet-18 and ViT-Tiny backbones, in in-domain and cross-domain settings, and proposes ECG-specific weak/strong augmentations and a fixed set of hyperparameters. The central claims are that semi-supervised methods improve delineation under label scarcity, that ViT-Tiny outperforms ResNet-18, and that the benchmark provides a standardized evaluation framework. Results are reported as mIoU and clinical interval MAE.

Significance. The benchmark resource itself is valuable: it unifies several underused datasets, standardizes preprocessing and evaluation, reports clinically relevant interval errors in addition to mIoU, and ships a public code repository. If the reported rankings were robust, the finding that ViT-Tiny benefits more from semi-supervised training than ResNet-18 would be an interesting and testable hypothesis for ECG delineation. The augmentation strategies and hyperparameter choices are useful starting points for the community. However, the empirical support for the headline comparative claims is currently limited by single-run reporting and configuration-transfer concerns, so the significance of the specific conclusions is conditional on additional experimental evidence.

major comments (4)
  1. [§3.3, Tables 2–4] All augmentation and SemiSeg hyperparameters were tuned once on LUDB at the 1/16 label ratio with a ResNet-18 backbone and then applied unchanged to every other database, label ratio, the ViT backbone, and the cross-domain setting. Table 2 already contains clear transfer failures: MT with ResNet-18 on QTDB at 1/8 drops to 44.8 mIoU versus 56.2 for Scratch, and ReCo with ViT-Tiny on Zhejiang at 1/2 collapses to 63.3 versus 80.2 for Scratch. Because methods are compared at a single fixed configuration, the headline claim that the transformer outperforms the convolutional network and the general claim that SemiSeg helps may reflect which architecture happens to suit the LUDB-tuned configuration rather than genuine method or architecture differences. Please provide per-method or per-dataset tuning results, or a sensitivity analysis demonstrating that the transferred hyperparameters are not responsible for the observed rankings.
  2. [Tables 2–4] All reported results are single-run point estimates; no standard deviations, confidence intervals, or significance tests are reported. The text draws conclusions from gaps as small as 0.1–0.2 mIoU points (e.g., LUDB ViT at 1/8: MT 76.7 vs. FixMatch 76.6) and uses them to rank methods. Given the stochasticity of deep learning training, these differences are within plausible noise. Please run at least three random seeds per condition and report mean ± standard deviation, or perform appropriate significance testing, and adjust the conclusions accordingly.
  3. [§4.2, Table 2] The statement that "the results confirm the effectiveness of SemiSeg algorithms on ECG delineation" is contradicted by multiple cells of Table 2, including ReCo with ViT-Tiny on Zhejiang at 1/2 (63.3 vs. 80.2 Scratch), MT with ResNet-18 on QTDB at 1/8 (44.8 vs. 56.2), and FixMatch with ResNet-18 on QTDB at 1/16 (46.7 vs. 47.5). The claim should be restricted to the settings where gains are consistent, and the failure cases should be discussed as part of the benchmark's value rather than dismissed.
  4. [§4.3, Table 4] The out-domain evaluation on mECGDB reports only interval MAE because the private database lacks delineation annotations, and mECGDB is not available to readers. The conclusion that ViT-Tiny "consistently benefited" from SemiSeg in the cross-domain setting while ResNet-18 did not therefore rests on a private, non-reproducible evaluation. Please either release de-identified mECGDB data or clearly mark the mECGDB results as preliminary and base the cross-domain claims primarily on the reproducible public-dataset portion of the benchmark.
minor comments (5)
  1. [Table 1] The column header "All Train Validation Test" is ambiguous; it would be clearer to present the split sizes as separate columns or with explicit spacing.
  2. [Figure 2] No error bars or number of runs are reported for the augmentation ablations, so differences among the strong augmentations (69.0–69.7 mIoU) may not be meaningful.
  3. [§2.2] The label set for QTDB and Zhejiang taken from reference [14] is not described; please state which annotations are used and how they differ from the original database labels.
  4. [§2.3] The text says results were obtained from checkpoints selected by highest validation mIoU; this selection criterion can bias the interval MAE comparisons. Please also report MAE at the checkpoint selected by MAE on the validation set, or justify the choice of a single selection metric.
  5. [§4.1] Minor wording issues include "confused the model's temporal cues" (Section 4.1) and inconsistent spacing around "PTB-XL" in Section 4.3; a careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the benchmark results are empirical measurements, with a mild augmentation-tuning caveat that does not make any claimed result equivalent to its inputs.

full rationale

SemiSegECG is an empirical benchmarking study rather than a derivation chain. The central claims—that SemiSeg methods generally improve over supervised training under label scarcity and that ViT-Tiny outperforms ResNet-18—are supported by held-out test-set measurements in Tables 2–4 and do not reduce to any fitted parameter by construction. The augmentation and SemiSeg hyperparameters were tuned on LUDB at the 1/16 label ratio (Sections 3.2–3.3) and then fixed across all other databases and settings, so the LUDB 1/16 row is mildly optimistic; this is an evaluation-leakage or correctness caveat, not circularity, because the cross-method comparisons and the cross-domain/ViT results are not derived from the tuning objective. The self-citations [21, 28] support generic background claims about ECG-specific augmentation and architecture reliability, but they are not load-bearing: the benchmark protocol, augmentations, and conclusions rest on the publicly available datasets and the reported experiments, not on an unverified prior result by the same authors. No equation is defined in terms of its target, no fitted value is renamed as a prediction, and no uniqueness or ansatz is imported from the authors' prior work. Accordingly, the paper is self-contained with respect to circularity concerns, and the score is 0.

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

The paper's central numbers depend on several modeling choices made in Sections 2.2-2.3: lead-level independence, resampling/filtering, external label quality, and the use of PTB-XL as unlabeled data. The tuned hyperparameters in Section 3.3 are free parameters selected on a single configuration. There are no invented physical entities.

free parameters (4)
  • FixMatch confidence threshold = 0.8
    Selected by hyperparameter tuning on LUDB at the 1/16 label ratio with ResNet-18 (Section 3.3); it controls which unlabeled examples contribute pseudo-labels and directly affects all FixMatch results.
  • ReCo confidence thresholds = easy 0.65, hard 0.8
    Chosen on the same LUDB 1/16 configuration (Section 3.3); defines the region of pixels used for contrastive queries, influencing ReCo results.
  • EMA decay = 0.99
    Set for Mean Teacher on LUDB 1/16 (Section 3.3); determines how quickly teacher weights track the student and shapes MT results.
  • Augmentation policy = random resized crop + four-operation RandAugment
    The weak and strong augmentation sets were selected empirically on LUDB 1/16 with ResNet (Scratch for weak, FixMatch for strong) and then applied to all methods and datasets (Section 3.2, Figure 2); this categorical choice is a tuned free parameter of the benchmark.
assumptions (4)
  • domain assumption Each ECG lead is treated as an independent training instance, and sample counts are lead-based rather than subject-based.
    Invoked in Section 2.2 to enlarge the training set and justify the reported sample counts; assumes lead-wise independence despite leads within a recording sharing a heartbeat and pathology.
  • domain assumption All signals are resampled to 250 Hz and bandpass filtered 0.67-40 Hz before training.
    Stated in Section 2.2 to uniformize resolution and remove noise; assumes downsampling from 500-2000 Hz sources and the filter do not remove clinically relevant boundary information used for P/QRS/T offsets.
  • domain assumption The QTDB and Zhejiang delineation labels released by a previous study [14] are accurate and complete enough for benchmarking.
    Used in Section 2.2 as ground truth for two of the four labeled databases; errors in these external labels directly enter all model scores.
  • domain assumption PTB-XL can serve as out-domain unlabeled data for ECG delineation even though it was collected for classification and has no wave-boundary labels.
    Stated in Section 2.1 and Section 2.3; assumes pseudo-labels generated on PTB-XL are informative rather than misleading for waveform boundaries.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SemiSegECG: A Multi-Dataset Benchmark for Semi-Supervised Semantic Segmentation in ECG Delineation." pith.science (2026). https://pith.science/paper/O2A6PFMW

@misc{pith2026250718323,
  author       = {Pith},
  title        = {Pith review of: SemiSegECG: A Multi-Dataset Benchmark for Semi-Supervised Semantic Segmentation in ECG Delineation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O2A6PFMW}},
  note         = {Machine review of arXiv:2507.18323}
}
read the original abstract

Electrocardiogram (ECG) delineation, the segmentation of meaningful waveform features, is critical for clinical diagnosis. Despite recent advances using deep learning, progress has been limited by the scarcity of publicly available annotated datasets. Semi-supervised learning presents a promising solution by leveraging abundant unlabeled ECG data. In this study, we present SemiSegECG, the first systematic benchmark for semi-supervised semantic segmentation (SemiSeg) in ECG delineation. We curated and unified multiple public datasets, including previously underused sources, to support robust and diverse evaluation. We adopted five representative SemiSeg algorithms from computer vision, implemented them on two different architectures: the convolutional network and the transformer, and evaluated them in two different settings: in-domain and cross-domain. Additionally, we propose ECG-specific training configurations and augmentation strategies and introduce a standardized evaluation framework. Our results show that the transformer outperforms the convolutional network in semi-supervised ECG delineation. We anticipate that SemiSegECG will serve as a foundation for advancing semi-supervised ECG delineation methods and will facilitate further research in this domain.

Figures

Figures reproduced from arXiv: 2507.18323 by the authors.

Figure 1
Figure 1. Examples of clean (left panel) and noisy (right panel) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of augmentation strategies. Left panel: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 38 canonical work pages

  1. [1]

    Isp ecg delineation dataset, 2024

    Aram Avetisyan, Nikolas Khachaturov, Ariana Asatryan, Shahane Tigranyan, and Yury Markin. Isp ecg delineation dataset, 2024. 2

  2. [2]

    Ecgvednet: A variational encoder-decoder network for ecg delineation in morphology variant ecgs

    Long Chen, Zheheng Jiang, Joseph Barker, Huiyu Zhou, Fer- nando Schlindwein, Will Nicolson, G Andre Ng, and Xin Li. Ecgvednet: A variational encoder-decoder network for ecg delineation in morphology variant ecgs. IEEE Transactions on Biomedical Engineering, 71(7):2143–2153, 2024. 1

  3. [3]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017. 2

  4. [4]

    Semi-supervised semantic segmentation with cross pseudo supervision

    Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 2613–2622, 2021. 2

  5. [5]

    Post-processing refined ecg de- lineation based on 1d-unet

    Zhenqin Chen, Mengying Wang, Meiyu Zhang, Wei Huang, Hanjie Gu, and Jinshan Xu. Post-processing refined ecg de- lineation based on 1d-unet. Biomedical Signal Processing and Control, 79:104106, 2023. 1, 2

  6. [6]

    Randaugment: Practical automated data augmen- tation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmen- tation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702–703, 2020. 3

  7. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representa- tions, 2021. 2

  8. [8]

    Revisiting qrs detection methodologies for portable, wearable, battery-operated, and wireless ecg sys- tems

    Mohamed Elgendi, Bj ¨orn Eskofier, Socrates Dokos, and Derek Abbott. Revisiting qrs detection methodologies for portable, wearable, battery-operated, and wireless ecg sys- tems. PloS one, 9(1):e84018, 2014. 1

Show all 44 references
  1. [9]

    ECG signal processing, classification and interpretation: a comprehensive frame- work of computational intelligence

    Adam Gacek and Witold Pedrycz. ECG signal processing, classification and interpretation: a comprehensive frame- work of computational intelligence . Springer Science & Business Media, 2011. 1, 2

  2. [10]

    Physiobank, physiotoolkit, and physionet: compo- nents of a new research resource for complex physiologic signals

    Ary L Goldberger, Luis AN Amaral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G Mark, Joseph E Mi- etus, George B Moody, Chung-Kang Peng, and H Eugene Stanley. Physiobank, physiotoolkit, and physionet: compo- nents of a new research resource for complex physiologi...

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2, 3

  4. [12]

    U-net architecture for the automatic detection and de- lineation of the electrocardiogram

    Guillermo Jimenez-Perez, Alejandro Alcaine, and Oscar Ca- mara. U-net architecture for the automatic detection and de- lineation of the electrocardiogram. In 2019 Computing in Cardiology (CinC), pages Page–1. IEEE, 2019. 1

  5. [13]

    Delineation of the electrocardiogram with a mixed- quality-annotations dataset using convolutional neural net- works

    Guillermo Jimenez-Perez, Alejandro Alcaine, and Oscar Ca- mara. Delineation of the electrocardiogram with a mixed- quality-annotations dataset using convolutional neural net- works. Scientific reports, 11(1):863, 2021. 2

  6. [14]

    Generalising electrocardiogram detec- tion and delineation: training convolutional neural networks with synthetic data augmentation

    Guillermo Jimenez-Perez, Juan Acosta, Alejandro Alcaine, and Oscar Camara. Generalising electrocardiogram detec- tion and delineation: training convolutional neural networks with synthetic data augmentation. Frontiers in Cardiovascu- lar Medicine, 11:1341786, 2024. 1, 2

  7. [15]

    Deep learning based ecg segmentation for delineation of diverse arrhyth- mias

    Chankyu Joung, Mijin Kim, Taejin Paik, Seong-Ho Kong, Seung-Young Oh, Won Kyeong Jeon, Jae-hu Jeon, Joong-Sik Hong, Wan-Joong Kim, Woong Kook, et al. Deep learning based ecg segmentation for delineation of diverse arrhyth- mias. PloS one, 19(6):e0303178, 2024. 1

  8. [16]

    Ludb: a new open-access validation tool for electrocar- diogram delineation algorithms

    Alena I Kalyakulina, Igor I Yusipov, Viktor A Moskalenko, Alexander V Nikolskiy, Konstantin A Kosonogov, Grigory V Osipov, Nikolai Yu Zolotykh, and Mikhail V Ivanchenko. Ludb: a new open-access validation tool for electrocar- diogram delineation algorithms. IEEE access , 8:186...

  9. [17]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. Advances in neural information processing systems , 25, 2012. 2

  10. [18]

    A database for evaluation of algorithms for mea- surement of qt and other waveform intervals in the ecg

    Pablo Laguna, Roger G Mark, A Goldberg, and George B Moody. A database for evaluation of algorithms for mea- surement of qt and other waveform intervals in the ecg. In Computers in cardiology 1997, pages 673–676. IEEE, 1997. 2

  11. [19]

    Efficient data augmentation policy for electrocardiograms

    Byeong Tak Lee, Yong-Yeon Jo, Seon-Yu Lim, Youngjae Song, and Joon-myoung Kwon. Efficient data augmentation policy for electrocardiograms. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 4153–4157, 2022. 3

  12. [20]

    Ecg segnet: An ecg delineation model based on the encoder-decoder structure

    Xiaohong Liang, Liping Li, Yuanyuan Liu, Dan Chen, Xin- pei Wang, Shunbo Hu, Jikuo Wang, Huan Zhang, Chengfa Sun, and Changchun Liu. Ecg segnet: An ecg delineation model based on the encoder-decoder structure. Computers in biology and medicine, 145:105445, 2022. 1, 2

  13. [21]

    Specialized ecg data augmentation method: leveraging precordial lead positional variability

    Jeonghwa Lim, Yeha Lee, Wonseuk Jang, and Sunghoon Joo. Specialized ecg data augmentation method: leveraging precordial lead positional variability. Biomedical Engineer- ing Letters, pages 1–12, 2025. 3

  14. [22]

    Performance analy- sis of ten common qrs detectors on different ecg application cases

    Feifei Liu, Chengyu Liu, Xinge Jiang, Zhimin Zhang, Yatao Zhang, Jianqing Li, and Shoushui Wei. Performance analy- sis of ten common qrs detectors on different ecg application cases. Journal of healthcare engineering, 2018(1):9050812,

  15. [23]

    Bootstrapping semantic segmentation with re- gional contrast

    Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew Davison. Bootstrapping semantic segmentation with re- gional contrast. In International Conference on Learning Representations, 2022. 3

  16. [24]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Pro- ceedings of the IEEE conference on computer vision and pat- tern recognition, pages 3431–3440, 2015. 2

  17. [25]

    SGDR: Stochastic gradi- ent descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradi- ent descent with warm restarts. In International Conference on Learning Representations, 2017. 3

  18. [26]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. In International Conference on Learning Representations, 2019. 3

  19. [27]

    A wavelet-based ecg delineator: evaluation on standard databases

    Juan Pablo Mart ´ınez, Rute Almeida, Salvador Olmos, Ana Paula Rocha, and Pablo Laguna. A wavelet-based ecg delineator: evaluation on standard databases. IEEE Transac- tions on biomedical engineering, 51(4):570–581, 2004. 1

  20. [28]

    Guiding masked representation learning to capture spatio- temporal relationship of electrocardiogram

    Yeongyeon Na, Minje Park, Yunwon Tae, and Sunghoon Joo. Guiding masked representation learning to capture spatio- temporal relationship of electrocardiogram. In International Conference on Learning Representations, 2024. 2

  21. [29]

    Data augmentation for electro- cardiogram classification with deep neural network

    Naoki Nonaka and Jun Seita. Data augmentation for electro- cardiogram classification with deep neural network. arXiv preprint arXiv:2009.04398, 2020. 3

  22. [30]

    A survey on semi-supervised semantic segmentation

    Adrian Pel ´aez-Vegas, Pablo Mesejo, and Juli ´an Luengo. A survey on semi-supervised semantic segmentation. arXiv preprint arXiv:2302.09899, 2023. 1

  23. [31]

    A systematic survey of data augmentation of ecg signals for ai applications

    Md Moklesur Rahman, Massimo Walter Rivolta, Fabio Badilini, and Roberto Sassi. A systematic survey of data augmentation of ecg signals for ai applications. Sensors, 23 (11):5237, 2023. 3

  24. [32]

    Automatic diagnosis of the 12-lead ecg using a deep neural network.Nature commu- nications, 11(1):1760, 2020

    Ant ˆonio H Ribeiro, Manoel Horta Ribeiro, Gabriela MM Paix˜ao, Derick M Oliveira, Paulo R Gomes, J ´essica A Canazart, Milton PS Ferreira, Carl R Andersson, Peter W Macfarlane, Wagner Meira Jr, et al. Automatic diagnosis of the 12-lead ecg using a deep neural network.Nature c...

  25. [33]

    Multiple electrocardiogram generator with single-lead electrocardiogram

    Hyo-Chang Seo, Gi-Won Yoon, Segyeong Joo, and Gi- Byoung Nam. Multiple electrocardiogram generator with single-lead electrocardiogram. Computer Methods and Pro- grams in Biomedicine, 221:106858, 2022. 2

  26. [34]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:5...

  27. [35]

    Dropout: a simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. 3

  28. [36]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2

  29. [37]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. In International conference on machine learning , pages 10347–10357. PMLR, 2021. 3

  30. [38]

    Normal values of corrected heart- rate variability in 10-second electrocardiograms for all ages

    Marten E Van den Berg, Peter R Rijnbeek, Maartje N Niemeijer, Albert Hofman, Gerard Van Herpen, Michiel L Bots, Hans Hillege, Cees A Swenne, Mark Eijgelsheim, Bruno H Stricker, et al. Normal values of corrected heart- rate variability in 10-second electrocardiograms for all ag...

  31. [39]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2

  32. [40]

    Ptb-xl, a large publicly available electrocardiog- raphy dataset

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Di- eter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiog- raphy dataset. Scientific data, 7(1):1–15, 2020. 2

  33. [41]

    St++: Make self-training work better for semi-supervised se- mantic segmentation

    Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised se- mantic segmentation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4268–4277, 2022. 3

  34. [42]

    Augmentation matters: A simple-yet- effective approach to semi-supervised semantic segmenta- tion

    Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet- effective approach to semi-supervised semantic segmenta- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 11350–11359,

  35. [43]

    A 12-lead ecg database to identify ori- gins of idiopathic ventricular arrhythmia containing 334 pa- tients

    Jianwei Zheng, Guohua Fu, Kyle Anderson, Huimin Chu, and Cyril Rakovski. A 12-lead ecg database to identify ori- gins of idiopathic ventricular arrhythmia containing 334 pa- tients. Scientific data, 7(1):98, 2020. 2

  36. [44]

    Rethinking semantic segmen- tation from a sequence-to-sequence perspective with trans- formers

    Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmen- tation from a sequence-to-sequence perspective with trans- formers. In Proceedings of the IEEE/CVF conference...

Pith tools

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