Pith. sign in

REVIEW 4 major objections 6 minor 45 references

S2M2ECG: Spatio-temporal bi-directional State Space Model Enabled Multi-branch Mamba for ECG

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A 705K-parameter Mamba model matches ECG classifiers

desk verdict A sensible Mamba-for-ECG architecture with honest limitations, but the comparison numbers need matched re-runs and Table 5 needs a fix before I'd trust the superiority claim. read the letter →

arxiv 2509.03066 v1 pith:VXVBHVV7 submitted 2025-09-03 eess.SP cs.AIcs.LG

Huaicheng Zhang , Ruoxin Wang , Chenlian Zhou , Jiguang Shi , Yue Ge , Zhoutong Li , Sheng Chang , Hao Wang
show 2 more authors
Jin He Qijun Huang
This is my paper
classification eess.SPcs.AIcs.LG
keywords ECGclassificationstatespacemodelsMambamulti-leadbi-directionalscanningleadfusionarrhythmiadetectionlightweightdeeplearning
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

This paper sets out to show that a state space model built specifically for 12-lead ECGs can get the accuracy of deep ECG classifiers without their parameter and compute cost. The proposed S2M2ECG splits each lead's signal into segments, runs each lead through its own bidirectional Mamba encoder, then fuses the lead streams with a cross-lead module. On the rhythm-focused Chapman set and the clinical SNPH set, the authors report the best accuracy and F1 among the CNNs, RNNs, Transformers, GNNs, and hybrids they compare; on PTB-XL it trails the best CNN by 0.012 F1 while using about 0.705M parameters. The aim is to establish that Mamba's linear-time, input-dependent state transitions fit the multi-lead spatio-temporal structure of ECG, making lightweight real-time deployment plausible. The paper itself acknowledges the morphology gap and says Mamba's global prioritization can miss fine local waveform detail.

What carries the argument

The architecture is S2M2ECG: a per-lead multi-branch design wrapped around bidirectional Mamba encoders. Segment tokenization cuts each lead into equal-length patches embedded as tokens, each lead receives its own Mamba encoder plus classification token, and bidirectional scanning processes the token sequence forward and backward before concatenating. The lead fusion module—a feed-forward network for temporal enrichment and a squeeze-and-excitation network for channel-wise lead attention—recombines the 12 streams before a MLP classification head. Together these parts give local receptive-field behavior, long-range rhythm context, and cross-lead spatial fusion in linear time.

What would settle it

Retrain S2M2ECG and the best Table 7 baselines with identical 250 Hz resampling, 9-level db6 wavelet filtering, Z-score normalization, segment length and tokenization, class balancing, and an equal hyperparameter-tuning budget. If S2M2ECG no longer leads on Chapman and SNPH, or its PTB-XL F1 gap exceeds 0.012, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that ECG classification becomes both more accurate and much cheaper if a 12-lead ECG is treated as sub-2D data: tokenize each lead's time series into segments, scan each lead's token sequence with a dedicated bidirectional Mamba encoder, and then fuse the 12 lead-specific feature streams with a temporal FFN plus spatial squeeze-and-excitation module. Bidirectional scanning supplies forward and backward context, segment tokens localize morphology, the multi-branch encoders preserve lead identity, and lead fusion reintegrates spatial dependencies. The authors report that this beats or ties every compared baseline on Chapman and SNPH, comes close to the best CNN on PTB-XL,

Load-bearing premise

The reported performance gap assumes the compared baselines were trained and tuned under the same preprocessing, input length, and per-dataset hyperparameter-selection conditions as S2M2ECG; the paper does not demonstrate this, and no code or training protocol is released.

Editorial extensions

If this is right

  • If the reported numbers hold, a single SSM architecture can serve rhythm diagnosis and clinical screening with around 0.705M parameters, roughly an order of magnitude smaller than CNN and Transformer baselines.
  • The measured CPU inference times of roughly 1–10 ms for a 10-second ECG imply that real-time analysis on resource-constrained or wearable devices is feasible.
  • Cross-database transfer between Chapman and NFH loses little accuracy, suggesting clinically trained SSM classifiers could generalize across hospital data sources.
  • Because scanning scales linearly with sequence length, long recordings such as 24-hour Holter data become tractable for the same architecture.
  • The PTB-XL results indicate morphology-heavy tasks may still favor CNN inductive bias, so the paper's own proposed CNN or Transformer hybrids are a natural next step.

Reading between the lines

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

  • My inference: the per-lead tokenization and bidirectional scanning could be adapted to single-lead wearable streams by dropping the fusion branch, a variant the paper does not test.
  • My inference: the paper's own future-work hint points to CNN-frontend hybrids as the most direct test of whether the PTB-XL morphology gap can be closed without losing the parameter savings.
  • My inference: because the per-dataset segment length, scan step, and feature dimension are chosen after seeing test results, the reported margins should be re-examined under a fixed training protocol before treating them as architecture-only gains.
  • My inference: token-level saliency or lead-attention visualizations, analogous to Transformer attention maps, would be a natural way to address the explainability limitation the paper acknowledges.
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 / 6 minor

Summary. The paper proposes S2M2ECG, a Mamba/SSM-based architecture for 12-lead ECG classification. The method combines per-lead segment tokenization, per-lead Mamba encoders with bidirectional scanning, and a lead-fusion module with FFN and SE attention, followed by an MLP classifier. Experiments are reported on four datasets (PTB-XL, Chapman, SNPH, NFH) covering morphological, rhythm, and clinical scenarios. The paper claims superior performance on Chapman and SNPH with among the fewest parameters (0.705M), competitive performance on PTB-XL, and includes ablations, confidence intervals, t-tests, and inference-latency measurements.

Significance. If the comparative results were obtained under a fully matched protocol, the paper would make a meaningful contribution: it would demonstrate that a lightweight SSM with per-lead encoders and bidirectional scanning can match or beat CNN/Transformer/GNN baselines on ECG classification, with linear complexity and deployable latency. The paper is above the typical experimental bar for this literature in that it reports ablations, confidence intervals, and two significance tests, and it evaluates on four datasets including a private clinical one. The central architectural idea is coherent and the efficiency claims are plausible. However, the validity of the headline comparative claims depends on experimental-equality assumptions that are not demonstrated in the manuscript.

major comments (4)
  1. [§4.7, Table 7] The central claim that S2M2ECG outperforms existing methods on Chapman and SNPH rests entirely on Table 7, yet the manuscript does not state whether the baseline entries were recomputed under the same protocol or transcribed from the original papers. The preprocessing fixed in §4.1 (250 Hz resampling, 9-level db6 wavelet filtering, Z-score) and the exact input length/tokenization are not shown to be applied to the baselines. Several baselines are prior works by the same group (e.g., ST-ReGE [14]) with different evaluation setups. Without matched splits, preprocessing, class balancing, and comparable hyperparameter tuning, the reported margins (e.g., +4.15% Acc on Chapman) cannot be attributed to the architecture.
  2. [§4.3.1–§4.3.3 and Table 7] The per-dataset best configurations (segment length p=25 vs 50, scanning step, feature dimension 48 vs 192) are selected after inspecting test-set F1/Acc in the ablation figures, and the same test set is then used to report the final Table 7 numbers. This is selection on the test set and inflates the reported performance and the significance results in §4.8. The authors should either fix one configuration across datasets or perform hyperparameter selection on the validation split only, and then report test results for the chosen configuration.
  3. [§4.5, Table 5] Table 5 is internally inconsistent with the text. On Chapman, removing the lead fusion module gives Accuracy 0.982 whereas the full model gives 0.913, while F1 and AUC improve with the module (0.895→0.918 and 0.975→0.985). The text states the lead fusion module brings 'all-around enhancement', which is contradicted by the Accuracy row. This suggests the ablation numbers are not produced from one fixed, reproducible protocol, or there is a typo. This needs to be resolved before the ablation claims can be evaluated.
  4. [§4.8] The significance analysis is presented only for F1 and is conditioned on the test-set-selected hyperparameters discussed above. The confidence interval formula in Eq. (10) uses a t-value but the degrees of freedom and the number of runs are not fully described, and Table 8 reports only the positive half-width ('+0.005'), not a symmetric interval. The p-values would be more informative if the comparison used a prespecified protocol and reported the full distribution for both S2M2ECG and the baseline.
minor comments (6)
  1. [Abstract and §4.7] The abstract claims 'superior performance in the rhythmic, morphological, and clinical scenarios', but §4.7 explicitly states that S2M2ECG does not outperform the best model on the morphological dataset PTB-XL (F1 gap 0.012). The wording should be aligned with the actual results.
  2. [Algorithm 1] The pseudo-code contains typos: 'Discritinization' should be 'Discretization', and the relation between the sequence-index variable l and the loop over L (the number of Mamba blocks) is unclear. Please clarify the notation and correct the typo.
  3. [§2.2] There is a grammatical error: 'They In semi-supervised learning, Semi-Mamba-UNet is proposed.' Please revise this sentence.
  4. [Eq. (10), Table 8] The confidence interval uses a plus sign only ('+0.005'). Use ± notation and state the number of runs and degrees of freedom explicitly.
  5. [§4.9] The memory utilization of 1063–1245 MB for a 0.705M-parameter model is surprisingly high and likely reflects the ONNX runtime framework rather than the model itself. Please state whether this is peak process memory and clarify that it is not model memory.
  6. [§4.4] In Table 4, the text says that S2M2ECG without multi-branch 'achieves a comprehensive increase' but the table shows the opposite; this appears to be a typo ('with' vs 'without').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: S2M2ECG is an empirical architecture evaluation; its reported results do not reduce by construction to fitted inputs or to a self-citation chain.

full rationale

The paper contains no derivation chain in which a predicted quantity is defined in terms of the thing it predicts. The central claims—superior F1/AUC on Chapman, SNPH, and PTB-XL and 0.705M parameters—are empirical measurements against external datasets, not consequences of definitions. The multi-branch design is adopted by explicit inspiration from the authors' prior MFB-CBRNN ([13]) and is compared against ST-ReGE ([14]); these self-citations motivate design choices but do not force the reported results. The wavelet preprocessing follows an external reference [42], and the real-time threshold is from an external reference [45]. The only near-circularity concern is §4.3, where per-dataset hyperparameters (segment length p, scan step s, feature dimension, depth) are selected after inspecting test-set F1/accuracy and the same test set is then reported in Table 7; this is a statistical validity/overfitting concern, not a by-construction reduction of the claim to its inputs. There are also internal inconsistencies (e.g., §4.5 claims the lead fusion module gives 'all-around enhancement' while Table 5 shows Chapman Accuracy drops from 0.982 without the module to 0.913 with it), which are correctness/reproducibility risks, not circularity. Because no step equates a purported prediction with fitted values or an unverified self-citation chain, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The paper's contribution is an engineering artifact (a specific architecture), not a theory or law. There are no invented particles or forces. The free parameters are the standard hyper-parameters of the model: segment length, scan step, feature dimension, and depth, selected per dataset via ablation. The main unstated burden is the selection of these hyper-parameters after seeing results, which is a form of model selection on the test distribution, and the assumption that the baseline comparisons are fair.

free parameters (4)
  • segment length p = 50 for Chapman, 25 for PTB-XL and SNPH
    Selected per dataset based on validation/test F1 (Fig. 5), a free configuration that changes the receptive field.
  • scanning step s = p/2 for Chapman and PTB-XL, p for SNPH
    Chosen per dataset as a multiple of segment length after evaluating F1 (Fig. 5).
  • feature dimension d = 192 for Chapman and SNPH, 48 for PTB-XL
    Chosen per dataset from ablation (Fig. 6).
  • Mamba block depth = 12 for all datasets
    Chosen from ablation (Table 3), mostly consistent across datasets.
assumptions (3)
  • domain assumption The 12 leads of an ECG can be treated as independent channels and fused later via a learned module.
    This is the design premise of the multi-branch architecture (Section 3.3), validated only by the ablation in Table 4.
  • domain assumption The Mamba selective SSM discretization (ZOH with A, B, C, Delta parameters) is a correct and sufficient backbone for ECG feature extraction.
    The paper builds on Mamba without re-deriving it (Section 3.1), treating it as an external, correct method.
  • domain assumption The SNPH private dataset is representative of clinical ECG practice and its labels are accurate.
    The paper relies on doctor-verified annotations and a specific recorder (Section 4.1), but the data is not released and its labeling protocol cannot be independently checked.
invented entities (1)
  • S2M2ECG architecture (per-lead Mamba encoders + lead fusion module + bidirectional scanning) independent evidence
    purpose: ECG classification with low parameter count and strong rhythm accuracy
    The architecture itself is an invented artifact, but it has independent evidence through its reported performance on external datasets (PTB-XL, Chapman, NFH) and is falsifiable by future replication.

how reviews work

0 comments
Cite this review

Pith. "Pith review of S2M2ECG: Spatio-temporal bi-directional State Space Model Enabled Multi-branch Mamba for ECG." pith.science (2026). https://pith.science/paper/VXVBHVV7

@misc{pith2026250903066,
  author       = {Pith},
  title        = {Pith review of: S2M2ECG: Spatio-temporal bi-directional State Space Model Enabled Multi-branch Mamba for ECG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VXVBHVV7}},
  note         = {Machine review of arXiv:2509.03066}
}
read the original abstract

As one of the most effective methods for cardiovascular disease (CVD) diagnosis, multi-lead Electrocardiogram (ECG) signals present a characteristic multi-sensor information fusion challenge that has been continuously researched in deep learning domains. Despite the numerous algorithms proposed with different DL architectures, maintaining a balance among performance, computational complexity, and multi-source ECG feature fusion remains challenging. Recently, state space models (SSMs), particularly Mamba, have demonstrated remarkable effectiveness across various fields. Their inherent design for high-efficiency computation and linear complexity makes them particularly suitable for low-dimensional data like ECGs. This work proposes S2M2ECG, an SSM architecture featuring three-level fusion mechanisms: (1) Spatio-temporal bi-directional SSMs with segment tokenization for low-level signal fusion, (2) Intra-lead temporal information fusion with bi-directional scanning to enhance recognition accuracy in both forward and backward directions, (3) Cross-lead feature interaction modules for spatial information fusion. To fully leverage the ECG-specific multi-lead mechanisms inherent in ECG signals, a multi-branch design and lead fusion modules are incorporated, enabling individual analysis of each lead while ensuring seamless integration with others. Experimental results reveal that S2M2ECG achieves superior performance in the rhythmic, morphological, and clinical scenarios. Moreover, its lightweight architecture ensures it has nearly the fewest parameters among existing models, making it highly suitable for efficient inference and convenient deployment. Collectively, S2M2ECG offers a promising alternative that strikes an excellent balance among performance, computational complexity, and ECG-specific characteristics, paving the way for high-performance, lightweight computations in CVD diagnosis.

Figures

Figures reproduced from arXiv: 2509.03066 by the authors.

Figure 1
Figure 1. The illustration of lead systems and process of ECG signals slicing. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The workflow of S2M2ECG, including segment tokenization and model processing. different leads can yield a comprehensive analysis of an ECG. Different leads contain diverse information, thus it’s unsuitable to directly aggregate the information of the whole ECG. A feasible approach is analyzing signals on each lead individually, and aggregating them finally. Therefore, in the segment tokenization stage, signals on di… view at source ↗
Figure 3
Figure 3. The architecture and data workflow of ECG-Mamba module in S [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The lead fusion module and classifying head of S [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Ablation results on segment length and scanning step. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Ablation results on feature dimensions. rather than the later ones. For language, such designation works because it represents the semantic analyzing process. However, in the domain of ECGs, information that is hidden in the prior and later curves are both matters. The…
Figure 7
Figure 7. Figure 7: Ablation results on bi-directional scanning designation. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Experiments on inference latency of S2M2ECG. the inference time is less than 5ms, except for the one with the least patch size and the scanning step. However, even the latency of the most time-consuming model fully satisfies the real-time rate requirement. According to…
Figure 9
Figure 9. Figure 9: Experiments on inference memory of S2M2ECG. 4.10 Clinical Implications of S2M2ECG The superior performance of S2M2ECG on clinical ECG database SNPH carries significant implications for real-world medical applications. Automatic detection in clinical practice may witnes…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 34 canonical work pages

  1. [14]

    St-rege: A novel spatial- temporal residual graph convolutional network for cvd

    Huaicheng Zhang, Wenhan Liu, Sheng Chang, Hao Wang, Jin He, and Qijun Huang. St-rege: A novel spatial- temporal residual graph convolutional network for cvd. IEEE Journal of Biomedical and Health Informatics , 2023

    work page 2023

  2. [1]

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

    Awni Y Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H Tison, Codie Bourn, Mintu P Turakhia, and Andrew Y Ng. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature medicine, 25(1):65–69, 2019

    2019

  3. [2]

    Automated multilabel diagnosis on electrocardiographic images and signals

    Veer Sangha, Bobak J Mortazavi, Adrian D Haimovich, Antônio H Ribeiro, Cynthia A Brandt, Daniel L Jacoby, Wade L Schulz, Harlan M Krumholz, Antonio Luiz P Ribeiro, and Rohan Khera. Automated multilabel diagnosis on electrocardiographic images and signals. Nature communications, 13(1):1583, 2022

    work page 2022

  4. [3]

    Attention is all you need

    A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017

    2017

  5. [4]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023

    arXiv 2023

  6. [5]

    Deep residual learning for image recognition

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

    2016

  7. [6]

    Deep neural networks generalization and fine-tuning for 12-lead ecg classification

    Aram Avetisyan, Shahane Tigranyan, Ariana Asatryan, Olga Mashkova, Sergey Skorik, Vladislav Ananev, and Yury Markin. Deep neural networks generalization and fine-tuning for 12-lead ecg classification. Biomedical Signal Processing and Control, 93:106160, 2024

    work page 2024

  8. [7]

    Lstm: A search space odyssey

    Klaus Greff, Rupesh K Srivastava, Jan Koutník, Bas R Steunebrink, and Jürgen Schmidhuber. Lstm: A search space odyssey. IEEE transactions on neural networks and learning systems, 28(10):2222–2232, 2016

    work page 2016

Show all 45 references
  1. [8]

    Multi-class arrhythmia detection from 12-lead varied-length ecg using attention-based time-incremental convolutional neural network

    Qihang Yao, Ruxin Wang, Xiaomao Fan, Jikui Liu, and Ye Li. Multi-class arrhythmia detection from 12-lead varied-length ecg using attention-based time-incremental convolutional neural network. Information Fusion, 53:174–182, 2020

    2020

  2. [9]

    Automatic cardiac arrhythmia classification using combination of deep residual network and bidirectional lstm

    Runnan He, Yang Liu, Kuanquan Wang, Na Zhao, Yongfeng Yuan, Qince Li, and Henggui Zhang. Automatic cardiac arrhythmia classification using combination of deep residual network and bidirectional lstm. IEEE access, 7:102119–102135, 2019

    2019

  3. [10]

    Automated ecg classification based on 1d deep learning network

    Chun-Yen Chen, Yan-Ting Lin, Shie-Jue Lee, Wei-Chung Tsai, Tien-Chi Huang, Yi-Hsueh Liu, Mu-Chun Cheng, and Chia-Yen Dai. Automated ecg classification based on 1d deep learning network. Methods, 202:127–135, 2022

    2022

  4. [11]

    Multilabel 12-lead ecg classification based on leadwise grouping multibranch network

    Xiaoyun Xie, Hui Liu, Da Chen, Minglei Shu, and Yinglong Wang. Multilabel 12-lead ecg classification based on leadwise grouping multibranch network. IEEE Transactions on Instrumentation and Measurement, 71:1–11, 2022

    2022

  5. [12]

    Msgformer: A multi-scale grid transformer network for 12-lead ecg arrhythmia detection

    Changqing Ji, Liyong Wang, Jing Qin, Lu Liu, Yue Han, and Zumin Wang. Msgformer: A multi-scale grid transformer network for 12-lead ecg arrhythmia detection. Biomedical Signal Processing and Control, 87:105499, 2024

    2024

  6. [13]

    Mfb-cbrnn: A hybrid network for mi detection using 12-lead ecgs

    Wenhan Liu, Fei Wang, Qijun Huang, Sheng Chang, Hao Wang, and Jin He. Mfb-cbrnn: A hybrid network for mi detection using 12-lead ecgs. IEEE journal of biomedical and health informatics, 24(2):503–514, 2019

    2019

  7. [15]

    Honest-ge: 2-step heuristic optimization and node-level embedding empower spatial-temporal graph model for ecg

    Huaicheng Zhang, Wenhan Liu, Deyu Luo, Jiguang Shi, Qianxi Guo, Yue Ge, Sheng Chang, Hao Wang, Jin He, and Qijun Huang. Honest-ge: 2-step heuristic optimization and node-level embedding empower spatial-temporal graph model for ecg. Information Sciences, 677:120941, 2024. 20 Ru...

    2024

  8. [16]

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

    Mohammed Guhdar, Abdulhakeem O Mohammed, and Ramadhan J Mstafa. Advanced deep learning framework for ecg arrhythmia classification using 1d-cnn with attention mechanism. Knowledge-Based Systems, page 113301, 2025

    2025

  9. [17]

    A survey of mamba

    Haohao Qu, Liangbo Ning, Rui An, Wenqi Fan, Tyler Derr, Hui Liu, Xin Xu, and Qing Li. A survey of mamba. arXiv preprint arXiv:2408.01129, 2024

    2024 arXiv

  10. [18]

    Mambabyte: Token-free selective state space model

    Junxiong Wang, Tushaar Gangavarapu, Jing Nathan Yan, and Alexander M Rush. Mambabyte: Token-free selective state space model. arXiv preprint arXiv:2401.13660, 2024

    2024 arXiv

  11. [19]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024

    2024 arXiv

  12. [20]

    Vmamba: Visual state space model 2024

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model 2024. arXiv preprint arXiv:2401.10166, 2024

    2024 arXiv

  13. [21]

    Pan-mamba: Effective pan-sharpening with state space model

    Xuanhua He, Ke Cao, Jie Zhang, Keyu Yan, Yingying Wang, Rui Li, Chengjun Xie, Danfeng Hong, and Man Zhou. Pan-mamba: Effective pan-sharpening with state space model. Information Fusion, 115:102779, 2025

    2025

  14. [22]

    Polyp-mamba: A hybrid multi-frequency perception gated selection network for polyp segmentation

    Xingguo Zhu, Wei Wang, Chen Zhang, and Haifeng Wang. Polyp-mamba: A hybrid multi-frequency perception gated selection network for polyp segmentation. Information Fusion, 115:102759, 2025

    2025

  15. [23]

    Semi-mamba-unet: Pixel-level contrastive and cross-supervised visual mamba-based unet for semi-supervised medical image segmentation

    Chao Ma and Ziyang Wang. Semi-mamba-unet: Pixel-level contrastive and cross-supervised visual mamba-based unet for semi-supervised medical image segmentation. Knowledge-Based Systems, 300:112203, 2024

    2024

  16. [24]

    Frequency decoupling enhancement and mamba depth extraction-based feature fusion in transformer hashing image retrieval

    Jiayi Chen, Shuli Cheng, Liejun Wang, Yongming Li, and Qiang Zou. Frequency decoupling enhancement and mamba depth extraction-based feature fusion in transformer hashing image retrieval. Knowledge-Based Systems, page 113036, 2025

    2025

  17. [25]

    Timemachine: A time series is worth 4 mambas for long-term forecasting

    Md Atik Ahamed and Qiang Cheng. Timemachine: A time series is worth 4 mambas for long-term forecasting. arXiv preprint arXiv:2403.09898, 2024

    2024 arXiv

  18. [26]

    Tscmamba: Mamba meets multi-view learning for time series classification

    Md Atik Ahamed and Qiang Cheng. Tscmamba: Mamba meets multi-view learning for time series classification. Information Fusion, page 103079, 2025

    2025

  19. [27]

    Mgcn: Mamba-integrated spa- tiotemporal graph convolutional network for long-term traffic forecasting.Knowledge-Based Systems, 309:112875, 2025

    Wenxie Lin, Zhe Zhang, Gang Ren, Yangzhen Zhao, Jingfeng Ma, and Qi Cao. Mgcn: Mamba-integrated spa- tiotemporal graph convolutional network for long-term traffic forecasting.Knowledge-Based Systems, 309:112875, 2025

    2025

  20. [28]

    St-mambasync: Complement the power of mamba and transformer fusion for less computational cost in spatial–temporal traffic forecasting

    Zhiqi Shao, Ze Wang, Xusheng Yao, Michael GH Bell, and Junbin Gao. St-mambasync: Complement the power of mamba and transformer fusion for less computational cost in spatial–temporal traffic forecasting. Information Fusion, 117:102872, 2025

    2025

  21. [29]

    Ecgmamba: Towards efficient ecg classification with bissm

    Yupeng Qiang, Xunde Dong, Xiuling Liu, Yang Yang, Yihai Fang, and Jianhong Dou. Ecgmamba: Towards efficient ecg classification with bissm. arXiv preprint arXiv:2406.10098, 2024

    2024 arXiv

  22. [30]

    Mambacapsule: Towards transparent cardiac disease diagnosis with electrocardiography using mamba capsule network

    Yinlong Xu, Xiaoqiang Liu, Zitai Kong, Yixuan Wu, Yue Wang, Yingzhou Lu, Honghao Gao, Jian Wu, and Hongxia Xu. Mambacapsule: Towards transparent cardiac disease diagnosis with electrocardiography using mamba capsule network. arXiv preprint arXiv:2407.20893, 2024

    2024 arXiv

  23. [31]

    Msecg: Incorporating mamba for robust and efficient ecg super-resolution

    Jie Lin, I Chiu, Kuan-Chen Wang, Kai-Chun Liu, Hsin-Min Wang, Ping-Cheng Yeh, and Yu Tsao. Msecg: Incorporating mamba for robust and efficient ecg super-resolution. arXiv preprint arXiv:2412.04861, 2024

    2024 arXiv

  24. [32]

    Mecg-e: Mamba-based ecg enhancer for baseline wander removal

    Kuo-Hsuan Hung, Kuan-Chen Wang, Kai-Chun Liu, Wei-Lun Chen, Xugang Lu, Yu Tsao, and Chii-Wann Lin. Mecg-e: Mamba-based ecg enhancer for baseline wander removal. In 2024 IEEE International Conference on Big Data (BigData), pages 6469–6475. IEEE, 2024

    2024

  25. [33]

    Diagnostic and prognostic value of ambulatory ecg (holter) monitoring in patients with coronary heart disease: a review

    C Michael Gibson, Lauren N Ciaglo, Matthew C Southard, Shaun Takao, Caitlin Harrigan, Jason Lewis, Jason Filopei, Michelle Lew, Sabina A Murphy, and Jacqueline Buros. Diagnostic and prognostic value of ambulatory ecg (holter) monitoring in patients with coronary heart disease:...

    2007

  26. [34]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018

    2018

  27. [35]

    Ptb-xl, a large publicly available electrocardiography dataset

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiography dataset. Scientific data, 7(1):154, 2020

    2020

  28. [36]

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

    Jianwei Zheng, Jianming Zhang, Sidy Danioko, Hai Yao, Hangyuan Guo, and Cyril Rakovski. A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients. Scientific data, 7(1):48, 2020. 21 Running Title for Header

    2020

  29. [37]

    Optimal multi-stage arrhythmia classification approach

    Jianwei Zheng, Huimin Chu, Daniele Struppa, Jianming Zhang, Sir Magdi Yacoub, Hesham El-Askary, Anthony Chang, Louis Ehwerhemuepha, Islam Abudayyeh, Alexander Barrett, et al. Optimal multi-stage arrhythmia classification approach. Scientific reports, 10(1):2898, 2020

    2020

  30. [38]

    Circular on the issuance of a methodology for ethical review of research in the life sciences and medicine involving human beings, 2023

    National Health and Wellness Commission of China. Circular on the issuance of a methodology for ethical review of research in the life sciences and medicine involving human beings, 2023. Accessed on 2025-05-15

    2023

  31. [39]

    Notice on further improving the ethical review of life sciences and medical research involving human beings, 2023

    Shanghai Municipal Health Commission. Notice on further improving the ethical review of life sciences and medical research involving human beings, 2023. Accessed on 2025-05-15

    2023

  32. [40]

    Sigxcl: A signal-image-graph cross-modal contrastive learning framework for cvd diagnosis based on internet of medical things

    Huaicheng Zhang, Wenhan Liu, Zhoutong Li, Jiguang Shi, Sheng Chang, Hao Wang, Jin He, and Qijun Huang. Sigxcl: A signal-image-graph cross-modal contrastive learning framework for cvd diagnosis based on internet of medical things. IEEE Internet of Things Journal, 2023

    2023

  33. [41]

    aecg-18 series digital electrocardiographs, 2025

    Nalong Health. aecg-18 series digital electrocardiographs, 2025. Accessed on 2025-05-19

    2025

  34. [42]

    Ecg beat classification using pca, lda, ica and discrete wavelet transform

    Roshan Joy Martis, U Rajendra Acharya, and Lim Choo Min. Ecg beat classification using pca, lda, ica and discrete wavelet transform. Biomedical Signal Processing and Control, 8(5):437–448, 2013

    2013

  35. [43]

    Pathological q waves in myocardial infarction in patients treated by primary pci

    Ronak Delewi, Georges IJff, Tim P van de Hoef, Alexander Hirsch, Lourens F Robbers, Robin Nijveldt, Anja M van der Laan, Pieter A van der Vleuten, Cees Lucas, Jan GP Tijssen, et al. Pathological q waves in myocardial infarction in patients treated by primary pci. JACC: Cardiov...

    2013

  36. [44]

    The pivotal role of ecg in cardiomyopathies

    Elisa Silvetti, Oreste Lanza, Fabiana Romeo, Annamaria Martino, Elisa Fedele, Chiara Lanzillo, Cinzia Crescenzi, Francesca Fanisio, and Leonardo Calò. The pivotal role of ecg in cardiomyopathies. Frontiers in Cardiovascular Medicine, 10:1178163, 2023

    2023

  37. [45]

    A deep learning approach to on-node sensor data analytics for mobile or wearable devices

    Daniele Ravi, Charence Wong, Benny Lo, and Guang-Zhong Yang. A deep learning approach to on-node sensor data analytics for mobile or wearable devices. IEEE journal of biomedical and health informatics, 21(1):56–64, 2016. 22

    2016

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.