arxiv: 2605.08199 · v1 · submitted 2026-05-05 · 📡 eess.SP · cs.LG

Recognition: 2 theorem links

· Lean Theorem

Domain-Adaptive Arrhythmia Classification Using a Hybrid Transformer on Wearable Heart Signals

Maedeh H. Toosi , Siamak Mohammadi

Authors on Pith no claims yet

Pith reviewed 2026-05-12 01:01 UTC · model grok-4.3

classification 📡 eess.SP cs.LG

keywords arrhythmia classificationdomain adaptationhybrid transformerwearable ECGheart rate variabilitymaximum mean discrepancygeneralizationECG signal processing

0 comments

The pith

A hybrid transformer aligns clinical ECG and wearable signal distributions to classify arrhythmias with 95% F1-macro on unseen data.

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

The paper aims to solve the problem of domain shift when moving arrhythmia classification models from clinical ECG recordings to wearable devices. It introduces a hybrid transformer architecture that processes both the raw ECG waveform for beat morphology and a set of heart rate variability features for rhythm analysis. Representation alignment using Maximum Mean Discrepancy minimizes the difference in feature distributions across domains, allowing effective training on public datasets and testing on new wearable data. This matters because it enables reliable, continuous heart monitoring in everyday settings where clinical training data does not match device recordings.

Core claim

The hybrid transformer model processes continuous ECG signals alongside seven heart rate variability features, where the raw signal path captures beat-level morphological patterns and the HRV path encodes rhythm regularity statistics. To address domain shifts, Maximum Mean Discrepancy aligns feature distributions between source clinical datasets and target wearable data. Trained on five public ECG datasets, the model achieves an F1-macro of 95% and balanced accuracy of 96.15% on unseen wearable device data, showing only a 2% drop in F1-macro from seen-domain performance.

What carries the argument

The hybrid transformer with dual paths for raw ECG morphology and HRV statistics, integrated with MMD-based domain alignment to reduce distribution shifts.

If this is right

Models can be trained on existing clinical datasets and deployed on wearables for real-time arrhythmia monitoring.
The combination of morphological and statistical features provides complementary information that enhances generalization.
Minimal degradation in performance supports the use of such systems in home and ambulatory environments.
Learning robust representations from multiple source datasets mitigates biases from individual data collections.

Where Pith is reading between the lines

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

If MMD alignment preserves class discriminability, the method may generalize to other signal types like photoplethysmography for heart monitoring.
Further reductions in domain gap could be achieved by incorporating patient-specific adaptation techniques.
Validation on larger and more varied wearable datasets would strengthen evidence for broad applicability.

Load-bearing premise

Aligning feature distributions via MMD between source clinical ECG datasets and target wearable data preserves discriminative power for arrhythmia classes without introducing artifacts or losing beat-level morphological cues.

What would settle it

A significant drop in classification performance on wearable data, such as F1-macro falling below 90%, or specific loss of ability to detect certain arrhythmias due to altered morphological features after alignment, would indicate the assumption does not hold.

Figures

Figures reproduced from arXiv: 2605.08199 by Maedeh H. Toosi, Siamak Mohammadi.

**Figure 2.** Figure 2: Comparison of ECG Signals for Different Arrhythmias: The left panel shows an atrial fibrillation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of the proposed model, a comprehensive illustration of the proposed transformer-based [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: KDE plots for Shannon Entropy and Normalized RMSSD before(a) and after preprocessing(b) [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: Violin plots for Mean, Standard Deviation (STD), Root Mean Square of Successive Differences [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 7.** Figure 7: Impact of SNR on model performance. The mean performance degradation (absolute AUC loss) is plotted against varying levels of additive white Gaussian noise. Error bars represent the standard deviation. The results show that the model is resilient to noise, with performance loss decreasing significantly as signal quality improves. domains previously seen by the model during training. The results in [PITH… view at source ↗

read the original abstract

Cardiovascular disease remains the leading cause of death globally, underscoring the need for effective, accessible monitoring solutions, particularly through wearable devices that enable continuous, real-time tracking of heart rhythms in home settings. However, deploying deep learning models trained on clinical electrocardiogram (ECG) datasets to wearable devices remains challenging, as differences in recording equipment, signal quality, and patient populations introduce domain shifts that degrade model performance. We propose a hybrid transformer model that processes continuous ECG signals alongside seven heart rate variability (HRV) features, where the raw signal path captures beat-level morphological patterns and the HRV path encodes rhythm regularity statistics, allowing the model to jointly leverage complementary information from both representations. To enhance the model's ability to generalize across domains, we employ representation learning techniques, including Maximum Mean Discrepancy (MMD), a non-parametric kernel-based metric that quantifies the distance between feature distributions of different domains, to align feature distributions between source and target domains, addressing the challenge of domain shifts between public datasets and wearable device data. By leveraging five public ECG datasets for training, the model learns robust, generalized representations that mitigate domain-specific biases. When tested on wearable device data with an unseen domain, the model achieved an F1-macro 95% and balanced accuracy of 96.15%. These results demonstrate minimal performance degradation, with only a 2% drop in F1-macro compared to seen-domain evaluation, highlighting the model's generalization capabilities and its potential for reliable, real-time heart monitoring applications in home and ambulatory settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The hybrid transformer fuses raw ECG with HRV features and applies MMD to cut domain shift from clinical to wearable data, reporting 95% F1-macro on unseen wearables with a 2% drop.

read the letter

The main takeaway is that this paper builds a hybrid transformer that takes both continuous ECG signals and seven HRV features, then uses MMD to align distributions between clinical training sets and wearable test data. It reports 95% F1-macro and 96.15% balanced accuracy on the unseen wearable recordings, with only a 2% drop from the seen-domain numbers. That combination of morphology and rhythm statistics plus explicit cross-domain testing is the concrete step forward here. Training across five public ECG datasets and evaluating on real wearable recordings shows a practical focus on deployment gaps rather than just another benchmark run. The hybrid input design is a reasonable way to capture both beat shape and interval regularity for arrhythmia tasks. The soft spots sit in the missing verification. The abstract states the performance numbers but gives no architecture diagram, fusion details, dataset sizes, class balance, training protocol, or ablation results. Without those, it is hard to tell whether MMD actually kept the class-discriminative morphological cues or whether the model simply carried over source-domain patterns that happened to work. The concern that marginal alignment could erase PVC versus normal distinctions if they correlate with domain shift is still open, since there are no class-conditional checks or before-after distribution plots. This work is aimed at people building real-time arrhythmia detectors for consumer devices. Readers who need a starting pipeline for domain-adaptive biosignal models will find the setup useful even if they have to fill in the gaps themselves. It deserves a serious referee because the problem is relevant and the reported numbers are high enough to warrant closer inspection. I would send it for review and ask specifically for the architecture, ablations, and evidence that adaptation preserved class separation.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a hybrid Transformer model for arrhythmia classification that jointly processes raw continuous ECG signals (to capture beat-level morphology) and seven HRV features (to encode rhythm statistics). It applies Maximum Mean Discrepancy (MMD) to align feature distributions between five public clinical ECG source datasets and an unseen wearable-device target domain. The model is reported to achieve 95% F1-macro and 96.15% balanced accuracy on the wearable test set, corresponding to only a 2% drop relative to seen-domain performance.

Significance. If the central claim holds, the work would demonstrate a practical route to deploying deep-learning arrhythmia detectors on consumer wearables by mitigating domain shift without requiring target-domain labels. The hybrid raw-signal-plus-HRV design and multi-dataset training are sensible engineering choices that could support continuous ambulatory monitoring and reduce the performance gap between clinical-grade and home-use recordings.

major comments (3)

[Methods (MMD alignment subsection)] Methods (MMD alignment subsection): the manuscript provides no class-conditional MMD, no per-class t-SNE or Wasserstein distances before/after alignment, and no ablation that removes MMD while retaining the identical hybrid backbone. Without these, it is impossible to confirm that marginal distribution alignment preserves the morphological cues (e.g., PVC vs. normal beat morphology) that distinguish arrhythmia classes rather than collapsing them.
[Results (performance tables and text)] Results (performance tables and text): the headline 95% F1-macro and 2% degradation figures are presented without dataset sizes, class counts or imbalance ratios for the five source datasets and the wearable target, training/validation protocol, hyper-parameters, or any statistical significance tests (e.g., confidence intervals or paired tests). These omissions are load-bearing for the generalization claim.
[Abstract and §4] Abstract and §4: the architecture description omits fusion details (early vs. late fusion of raw-signal and HRV paths inside the Transformer), number of attention heads/layers, and input segmentation strategy, all of which are required to reproduce or interpret the reported numbers.

minor comments (2)

[Abstract] The abstract phrase 'F1-macro 95%' should read 'an F1-macro of 95%' for grammatical clarity.
[Figures] Figure captions and axis labels should explicitly state whether the reported metrics are macro-averaged and whether error bars represent standard deviation across folds or runs.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and insightful comments on our manuscript. We believe these suggestions will significantly improve the clarity and completeness of our work. Below, we provide point-by-point responses to the major comments.

read point-by-point responses

Referee: Methods (MMD alignment subsection): the manuscript provides no class-conditional MMD, no per-class t-SNE or Wasserstein distances before/after alignment, and no ablation that removes MMD while retaining the identical hybrid backbone. Without these, it is impossible to confirm that marginal distribution alignment preserves the morphological cues (e.g., PVC vs. normal beat morphology) that distinguish arrhythmia classes rather than collapsing them.

Authors: We agree that additional analyses would help confirm the benefits of MMD. Since the target domain is unlabeled, class-conditional MMD is not directly applicable. However, we have added an ablation study removing the MMD component while keeping the hybrid backbone, t-SNE visualizations of the feature distributions before and after alignment (using source labels for coloring), and Wasserstein distance metrics to the revised Methods and Results sections. These additions demonstrate that the alignment improves domain generalization without collapsing class distinctions. revision: yes
Referee: Results (performance tables and text): the headline 95% F1-macro and 2% degradation figures are presented without dataset sizes, class counts or imbalance ratios for the five source datasets and the wearable target, training/validation protocol, hyper-parameters, or any statistical significance tests (e.g., confidence intervals or paired tests). These omissions are load-bearing for the generalization claim.

Authors: We have substantially expanded the Results section to include all requested information: dataset sizes and class imbalance ratios for the five source datasets and the wearable target, the training and validation protocol, complete hyperparameter settings, and statistical significance with confidence intervals computed over multiple runs. A new supplementary table has been added to present these details clearly, reinforcing the validity of the reported 95% F1-macro and minimal degradation. revision: yes
Referee: Abstract and §4: the architecture description omits fusion details (early vs. late fusion of raw-signal and HRV paths inside the Transformer), number of attention heads/layers, and input segmentation strategy, all of which are required to reproduce or interpret the reported numbers.

Authors: We have revised both the Abstract and Section 4 to include the missing architectural specifications. The model employs late fusion by concatenating the outputs of the raw ECG signal Transformer and the HRV feature MLP before the classification layer. It consists of 6 layers with 8 attention heads, and processes the ECG signals in 10-second segments with overlapping windows. These updates ensure the work is fully reproducible. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical evaluation on held-out wearable data is independent of training inputs

full rationale

The paper trains a hybrid transformer on five public ECG datasets using MMD for domain alignment and reports F1-macro 95% and balanced accuracy 96.15% on separate unseen wearable recordings. These metrics arise from standard supervised training plus held-out testing; they are not algebraically forced by the model equations, nor do they reduce to fitted parameters by construction. No self-citation chain, uniqueness theorem, or ansatz is invoked to derive the performance numbers. The derivation chain consists of architecture definition, loss minimization, and empirical measurement on disjoint data, all of which remain falsifiable and non-tautological.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that MMD alignment transfers discriminative arrhythmia information across domains and that the chosen seven HRV features plus raw signal morphology are jointly sufficient. No new entities are postulated; the work uses standard neural components and a known discrepancy metric.

free parameters (1)

Selection of seven HRV features
The paper specifies exactly seven heart rate variability features to encode rhythm regularity; the choice is presented without derivation from data or theory.

axioms (1)

domain assumption MMD alignment preserves class-discriminative information while reducing domain shift
Invoked when stating that feature distribution alignment enables generalization to unseen wearable data without performance loss.

pith-pipeline@v0.9.0 · 5583 in / 1370 out tokens · 58078 ms · 2026-05-12T01:01:16.514510+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We employ representation learning techniques, including Maximum Mean Discrepancy (MMD), a non-parametric kernel-based metric that quantifies the distance between feature distributions of different domains, to align feature distributions between source and target domains
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

hybrid transformer model that processes continuous ECG signals alongside seven heart rate variability (HRV) features

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

70 extracted references · 70 canonical work pages · 1 internal anchor

[1]

W. H. Organization, Global report on hypertension: the race against a silent killer, World Health Organization, 2023. 15

work page 2023
[2]

P.Joseph, D.Leong, M.McKee, S.S.Anand, J.-D.Schwalm, K.Teo, A.Mente, S.Yusuf, Reducing the global burden of cardiovascular disease, part 1: the epidemiology and risk factors, Circulation research 121 (6) (2017) 677–694

work page 2017
[3]

J. K. Grant, C. E. Ndumele, S. S. Martin, The evolving landscape of cardiovascular risk assessment, JAMA 332 (12) (2024) 967–969

work page 2024
[4]

Quer, Prediction of atrial fibrillation from at-home single-lead ecg signals without arrhythmias, NPJ Digital Medicine 6 (1) (2023) 229

M.Gadaleta, P.Harrington, E.Barnhill, E.Hytopoulos, M.P.Turakhia, S.R.Steinhubl, G. Quer, Prediction of atrial fibrillation from at-home single-lead ecg signals without arrhythmias, NPJ Digital Medicine 6 (1) (2023) 229

work page 2023
[5]

Biton, M

S. Biton, M. Aldhafeeri, E. Marcusohn, K. Tsutsui, T. Szwagier, A. Elias, J. Oster, J. M. Sellal, M. Suleiman, J. A. Behar, Generalizable and robust deep learning algorithm for atrial fibrillation diagnosis across geography, ages and sexes, NPJ Digital Medicine 6 (1) (2023) 44

work page 2023
[6]

G. J. Williams, A. Al-Baraikan, F. E. Rademakers, F. Ciravegna, F. N. van de Vosse, A. Lawrie, A. Rothman, E. A. Ashley, M. R. Wilkins, P. V. Lawford, et al., Wearable technology and the cardiovascular system: the future of patient assessment, The Lancet Digital Health 5 (7) (2023) e467–e476

work page 2023
[7]

Rajpurkar, E

P. Rajpurkar, E. Chen, O. Banerjee, E. J. Topol, Ai in health and medicine, Nature medicine 28 (1) (2022) 31–38

work page 2022
[8]

Gerke, T

S. Gerke, T. Minssen, G. Cohen, Ethical and legal challenges of artificial intelligence- driven healthcare, in: Artificial intelligence in healthcare, Elsevier, 2020, pp. 295–336

work page 2020
[9]

D. M. Hawkins, The problem of overfitting, Journal of chemical information and com- puter sciences 44 (1) (2004) 1–12

work page 2004
[10]

C.-W. Park, S. W. Seo, N. Kang, B. Ko, B. W. Choi, C. M. Park, D. K. Chang, H. Kim, H. Kim, H. Lee, et al., Artificial intelligence in health care: Current applications and issues, Journal of Korean medical science 35 (42) (2020)

work page 2020
[11]

T. A. Lasko, E. V. Strobl, W. W. Stead, Why do probabilistic clinical models fail to transport between sites, npj Digital Medicine 7 (1) (2024) 53

work page 2024
[12]

Bachtiger, C

P. Bachtiger, C. F. Petri, F. E. Scott, S. R. Park, M. A. Kelshiker, H. K. Sahemey, B. Dumea, R. Alquero, P. S. Padam, I. R. Hatrick, et al., Point-of-care screening for heart failure with reduced ejection fraction using artificial intelligence during ecg- enabled stethoscope examination in london, uk: a prospective, observational, multicen- tre study, Th...

work page 2022
[13]

Ben-David, J

S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, J. W. Vaughan, A theory of learning from different domains, Machine learning 79 (2010) 151–175

work page 2010
[14]

G. M. Marcus, Evaluation and management of premature ventricular complexes, Cir- culation 141 (17) (2020) 1404–1418. 16

work page 2020
[15]

with the Special Contribution of the European Heart Rhythm Association (EHRA), E

D. with the Special Contribution of the European Heart Rhythm Association (EHRA), E. by the European Association for Cardio-Thoracic Surgery (EACTS), A. F. Mem- bers, A. J. Camm, P. Kirchhof, G. Y. Lip, U. Schotten, I. Savelieva, S. Ernst, I. C. Van Gelder, et al., Guidelines for the management of atrial fibrillation: the task force for the management of ...

work page 2010
[16]

J. Oh, S. Kim, N. Ho, J.-H. Kim, H. Song, S.-Y. Yun, Understanding cross-domain few- shot learning based on domain similarity and few-shot difficulty, Advances in Neural Information Processing Systems 35 (2022) 2622–2636

work page 2022
[17]

X. Song, S. Zheng, W. Cao, J. Yu, J. Bian, Efficient and effective multi-task grouping via meta learning on task combinations, Advances in Neural Information Processing Systems 35 (2022) 37647–37659

work page 2022
[18]

Ng, M.-T

Y. Ng, M.-T. Liao, T.-L. Chen, C.-K. Lee, C.-Y. Chou, W. Wang, Few-shot transfer learningforpersonalizedatrialfibrillationdetectionusingpatient-basedsiamesenetwork with single-lead ecg records, Artificial Intelligence in Medicine 144 (2023) 102644

work page 2023
[19]

Amirshahi, M

A. Amirshahi, M. H. Toosi, S. Mohammadi, S. Albini, P. D. Schiavone, G. Ansaloni, A. Aminifar, D. Atienza, Metawears: A shortcut in wearable systems lifecycle with only a few shots, arXiv preprint arXiv:2408.01988 (2024)

work page arXiv 2024
[20]

H. Li, S. J. Pan, S. Wang, A. C. Kot, Domain generalization with adversarial fea- ture learning, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5400–5409

work page 2018
[21]

Ganin, E

Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. March, V.Lempitsky, Domain-adversarialtrainingofneuralnetworks, Journalofmachinelearn- ing research 17 (59) (2016) 1–35

work page 2016
[22]

Tzeng, J

E. Tzeng, J. Hoffman, K. Saenko, T. Darrell, Adversarial discriminative domain adap- tation, in: Proceedings of the IEEE conference on computer vision and pattern recog- nition, 2017, pp. 7167–7176

work page 2017
[23]

J. Lai, H. Tan, J. Wang, L. Ji, J. Guo, B. Han, Y. Shi, Q. Feng, W. Yang, Practical intelligent diagnostic algorithm for wearable 12-lead ecg via self-supervised learning on large-scale dataset, Nature Communications 14 (1) (2023) 3741

work page 2023
[24]

X. Yue, Z. Zheng, S. Zhang, Y. Gao, T. Darrell, K. Keutzer, A. S. Vincentelli, Prototyp- ical cross-domain self-supervised learning for few-shot unsupervised domain adaptation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2021, pp. 13834–13844

work page 2021
[25]

Achituve, H

I. Achituve, H. Maron, G. Chechik, Self-supervised learning for domain adaptation on point clouds, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 123–133. 17

work page 2021
[26]

Ragab, E

M. Ragab, E. Eldele, Z. Chen, M. Wu, C.-K. Kwoh, X. Li, Self-supervised autoregressive domain adaptation for time series data, IEEE Transactions on Neural Networks and Learning Systems 35 (1) (2022) 1341–1351

work page 2022
[27]

F. M. Carlucci, A. D’Innocente, S. Bucci, B. Caputo, T. Tommasi, Domain gener- alization by solving jigsaw puzzles, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2229–2238

work page 2019
[28]

Tobin, R

J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, P. Abbeel, Domain randomiza- tion for transferring deep neural networks from simulation to the real world, in: 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), IEEE, 2017, pp. 23–30

work page 2017
[29]

Gretton, D

A. Gretton, D. Sejdinovic, H. Strathmann, S. Balakrishnan, M. Pontil, K. Fukumizu, B. K. Sriperumbudur, Optimal kernel choice for large-scale two-sample tests, Advances in neural information processing systems 25 (2012)

work page 2012
[30]

J. Zhou, B. Hu, W. Feng, Z. Zhang, X. Fu, H. Shao, H. Wang, L. Jin, S. Ai, Y. Ji, An ensemble deep learning model for risk stratification of invasive lung adenocarcinoma using thin-slice ct, NPJ Digital Medicine 6 (1) (2023) 119

work page 2023
[31]

Placido, B

D. Placido, B. Yuan, J. X. Hjaltelin, C. Zheng, A. D. Haue, P. J. Chmura, C. Yuan, J. Kim, R. Umeton, G. Antell, et al., A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories, Nature medicine 29 (5) (2023) 1113–1122

work page 2023
[32]

Zhang, X

K. Zhang, X. Liu, J. Xu, J. Yuan, W. Cai, T. Chen, K. Wang, Y. Gao, S. Nie, X. Xu, et al., Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images, Nature biomedical engineering 5 (6) (2021) 533–545

work page 2021
[33]

R.Arnaout, L.Curran, Y.Zhao, J.C.Levine, E.Chinn, A.J.Moon-Grady, Anensemble of neural networks provides expert-level prenatal detection of complex congenital heart disease, Nature medicine 27 (5) (2021) 882–891

work page 2021
[34]

G. Wang, M. Chen, Z. Ding, J. Li, H. Yang, P. Zhang, Inter-patient ecg arrhythmia heartbeat classification based on unsupervised domain adaptation, Neurocomputing 454 (2021) 339–349

work page 2021
[35]

M. Chen, G. Wang, Z. Ding, J. Li, H. Yang, Unsupervised domain adaptation for ecg arrhythmia classification, in: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), IEEE, 2020, pp. 304–307

work page 2020
[36]

T.Zhu, C.Uduku, K.Li, P.Herrero, N.Oliver, P.Georgiou, Enhancingself-management in type 1 diabetes with wearables and deep learning, npj Digital Medicine 5 (1) (2022) 78

work page 2022
[37]

C. Wang, Y. Li, Y. Tsuboshita, T. Sakurai, T. Goto, H. Yamaguchi, Y. Yamashita, A. Sekiguchi, H. Tachimori, A. D. N. I. http://orcid. org/0000-0001-5252-1965 Wang 18 Caihua 7 http://orcid. org/0000-0003-2490-4867 Li Yuanzhong 7 http://orcid. org/0000 0002-4814-8093 Goto Tsubasa 7, A high-generalizability machine learning framework for predictingtheprogres...

work page 1965
[38]

H. Cui, D. Wei, K. Ma, S. Gu, Y. Zheng, A unified framework for generalized low-shot medical image segmentation with scarce data, IEEE Transactions on Medical Imaging 40 (10) (2020) 2656–2671

work page 2020
[39]

X. Liu, Z. Jiang, J. Fromm, X. Xu, S. Patel, D. McDuff, Metaphys: few-shot adaptation for non-contact physiological measurement, in: Proceedings of the conference on health, inference, and learning, 2021, pp. 154–163

work page 2021
[40]

Gupta, S

P. Gupta, S. Bhaskarpandit, M. Gupta, Similarity learning based few shot learning for ecg time series classification, in: 2021 Digital Image Computing: Techniques and Applications (DICTA), IEEE, 2021, pp. 1–8

work page 2021
[41]

Z. Li, H. Wang, X. Liu, A one-dimensional siamese few-shot learning approach for ecg classification under limited data, in: 2021 43rd annual international conference of the IEEE engineering in medicine & biology society (EMBC), IEEE, 2021, pp. 455–458

work page 2021
[42]

W. Chen, T. Banerjee, E. John, A meta-transfer learning approach to ecg arrhythmia detection, in: 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), IEEE, 2022, pp. 1300–1305

work page 2022
[43]

T. Ai, Z. Yang, H. Hou, C. Zhan, C. Chen, W. Lv, Q. Tao, Z. Sun, L. Xia, Correlation of chest ct and rt-pcr testing for coronavirus disease 2019 (covid-19) in china: a report of 1014 cases, Radiology 296 (2) (2020) E32–E40

work page 2019
[44]

M. Lu, Y. Pan, D. Nie, F. Liu, F. Shi, Y. Xia, D. Shen, Smile: Sparse-attention based multiple instance contrastive learning for glioma sub-type classification using patholog- ical images, in: MICCAI Workshop on Computational Pathology, PMLR, 2021, pp. 159–169

work page 2021
[45]

Shome, T

D. Shome, T. Kar, S. N. Mohanty, P. Tiwari, K. Muhammad, A. AlTameem, Y. Zhang, A. K. J. Saudagar, Covid-transformer: Interpretable covid-19 detection using vision transformer for healthcare, International Journal of Environmental Research and Public Health 18 (21) (2021) 11086

work page 2021
[46]

J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, Y. Zhou, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306 (2021)

work page internal anchor Pith review Pith/arXiv arXiv 2021
[47]

Hatamizadeh, Y

A. Hatamizadeh, Y. Tang, V. Nath, D. Yang, A. Myronenko, B. Landman, H. R. Roth, D. Xu, Unetr: Transformers for 3d medical image segmentation, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 574–584. 19

work page 2022
[48]

Karimi, S

D. Karimi, S. D. Vasylechko, A. Gholipour, Convolution-free medical image seg- mentation using transformers, in: Medical Image Computing and Computer As- sisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part I 24, Springer, 2021, pp. 78–88

work page 2021
[49]

R. Hu, J. Chen, L. Zhou, A transformer-based deep neural network for arrhythmia detection using continuous ecg signals, Computers in Biology and Medicine 144 (2022) 105325

work page 2022
[50]

Pratiher, A

S. Pratiher, A. Srivastava, Y. B. Priyatha, N. Ghosh, A. Patra, A dilated residual vision transformer for atrial fibrillation detection from stacked time-frequency ecg representa- tions, in: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2022, pp. 1121–1125

work page 2022
[51]

C. Che, P. Zhang, M. Zhu, Y. Qu, B. Jin, Constrained transformer network for ecg signal processing and arrhythmia classification, BMC Medical Informatics and Decision Making 21 (1) (2021) 184

work page 2021
[52]

Z. I. Attia, P. A. Noseworthy, F. Lopez-Jimenez, S. J. Asirvatham, A. J. Deshmukh, B. J. Gersh, R. E. Carter, X. Yao, A. A. Rabinstein, B. J. Erickson, et al., An artificial intelligence-enabled ecg algorithm for the identification of patients with atrial fibrilla- tion during sinus rhythm: a retrospective analysis of outcome prediction, The Lancet 394 (1...

work page 2019
[53]

Sangha, B

V. Sangha, B. J. Mortazavi, A. D. Haimovich, A. H. Ribeiro, C. A. Brandt, D. L. Jacoby, W. L. Schulz, H. M. Krumholz, A. L. P. Ribeiro, R. Khera, Automated multilabel diagnosis on electrocardiographic images and signals, Nature communications 13 (1) (2022) 1583

work page 2022
[54]

Khurshid, S

S. Khurshid, S. Friedman, C. Reeder, P. Di Achille, N. Diamant, P. Singh, L. X. Har- rington, X. Wang, M. A. Al-Alusi, G. Sarma, et al., Ecg-based deep learning and clinical risk factors to predict atrial fibrillation, Circulation 145 (2) (2022) 122–133

work page 2022
[55]

Biton, S

S. Biton, S. Gendelman, A. H. Ribeiro, G. Miana, C. Moreira, A. L. P. Ribeiro, J. A. Behar, Atrial fibrillation risk prediction from the 12-lead electrocardiogram using digital biomarkers and deep representation learning, European Heart Journal-Digital Health 2 (4) (2021) 576–585

work page 2021
[56]

Clifford, C

G. Clifford, C. Liu, B. Moody, L.-W. Lehman, I. Silva, Q. Li, A. Johnson, R. Mark, Af classification from a short single lead ecg recording: the physionet computing in cardiology challenge 2017, Computing in Cardiology (Rennes: IEEE) 14 (2017)

work page 2017
[57]

X. Wan, Y. Liu, X. Mei, J. Ye, C. Zeng, Y. Chen, A novel atrial fibrillation auto- matic detection algorithm based on ensemble learning and multi-feature discrimination, Medical & Biological Engineering & Computing 62 (6) (2024) 1809–1820

work page 2024
[58]

M. S. Jahan, M. Mansourvar, S. Puthusserypady, U. K. Wiil, A. Peimankar, Short-term atrial fibrillation detection using electrocardiograms: A comparison of machine learning approaches, International Journal of Medical Informatics 163 (2022) 104790. 20

work page 2022
[59]

C. Xie, L. McCullum, A. Johnson, T. Pollard, B. Gow, B. Moody, Waveform database software package (wfdb) for python, PhysioNet (2022)

work page 2022
[60]

J. Pan, W. J. Tompkins, A real-time qrs detection algorithm, IEEE transactions on biomedical engineering (3) (1985) 230–236

work page 1985
[61]

Vaswani, Attention is all you need, Advances in Neural Information Processing Sys- tems (2017)

A. Vaswani, Attention is all you need, Advances in Neural Information Processing Sys- tems (2017)

work page 2017
[62]

G. B. Moody, R. G. Mark, The impact of the mit-bih arrhythmia database, IEEE engineering in medicine and biology magazine 20 (3) (2001) 45–50

work page 2001
[63]

S. D. Greenwald, R. S. Patil, R. G. Mark, Improved detection and classification of arrhythmias in noise-corrupted electrocardiograms using contextual information, IEEE, 1990

work page 1990
[64]

Moody, A new method for detecting atrial fibrillation using rr intervals, Proc

G. Moody, A new method for detecting atrial fibrillation using rr intervals, Proc. Com- put. Cardiol. 10 (1983) 227–230

work page 1983
[65]

Zheng, H

J. Zheng, H. Guo, H. Chu, A large scale 12-lead electrocardiogram database for arrhyth- mia study (version 1.0. 0), PhysioNet 2022Available online httpphysionet orgcontentecg arrhythmia10 0accessed on 23 (2022)

work page 2022
[66]

E. H. Hafshejani, R. Rabbani, Z. Azizi, M. Barekatain, E. Khoram, A. Fotowat-Ahmady, Ultra low-power system for remote ecg monitoring, in: 2021 28th National and 6th International Iranian Conference on Biomedical Engineering (ICBME), IEEE, 2021, pp. 1–8

work page 2021
[67]

S. Kim, S. Chon, J.-K. Kim, J. Kim, Y. Gil, S. Jung, Lightweight convolutional neu- ral network for real-time arrhythmia classification on low-power wearable electrocar- diograph, in: 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), IEEE, 2022, pp. 1915–1918

work page 2022
[68]

A.H.Ribeiro, M.H.Ribeiro, G.M.Paixão, D.M.Oliveira, P.R.Gomes, J.A.Canazart, M. P. Ferreira, C. R. Andersson, P. W. Macfarlane, W. Meira Jr, et al., Automatic diagnosis of the 12-lead ecg using a deep neural network, Nature communications 11 (1) (2020) 1760

work page 2020
[69]

Phukan, M

N. Phukan, M. S. Manikandan, R. B. Pachori, Afibri-net: A lightweight convolution neural network based atrial fibrillation detector, IEEE Transactions on Circuits and Systems I: Regular Papers 70 (12) (2023) 4962–4974

work page 2023
[70]

S. U. Hassan, M. S. Mohd Zahid, T. A. Abdullah, K. Husain, Classification of cardiac arrhythmia using a convolutional neural network and bi-directional long short-term memory, Digital health 8 (2022) 20552076221102766. 21

work page 2022