Pith. sign in

REVIEW 3 major objections 4 minor 56 references

MedGNN: Towards Multi-resolution Spatiotemporal Graph Learning for Medical Time Series Classification

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read MedGNN models EEG and ECG channels as multi-resolution graphs, applies attention to signal differences to cancel baseline drift, and reports top results against ten baselines on five datasets.

desk verdict Solid empirical paper with a genuinely new difference-attention mechanism for baseline wander; the margins are plausible but baseline tuning is under-documented, so read the '22 top-1' claim with caution. read the letter →

arxiv 2502.04515 v1 pith:CFQBNRY3 submitted 2025-02-06 cs.LG cs.AI

Wei Fan , Jingru Fei , Dingyu Guo , Kun Yi , Xiaozhuang Song , Haolong Xiang , Hangting Ye , Min Li This is my paper · ORCID
classification cs.LGcs.AI MSC 68T0768T10
keywords medicaltimeseriesclassificationspatiotemporalgraphlearningmulti-resolutiongraphsdifferenceattentionfrequencyconvolutionbaselinewanderEEGECG
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 proposes MedGNN, a framework for classifying multivariate medical time series such as EEG and ECG, and argues that existing models miss three things: channel relationships that change with the observation scale, slow baseline drifts that obscure meaningful fluctuations, and the complementary frequency-domain view of the signal. MedGNN addresses these with learnable channel graphs built at several temporal resolutions, self-attention applied to the first-order temporal difference of the signal, and a frequency-domain convolution branch, all fused by a multi-resolution graph transformer. The paper reports that MedGNN outperforms ten existing transformer and graph baselines: the best score on all five metrics in sample-based ADFD evaluation, and 22 top-1 plus 2 top-2 placements across the 25 dataset-metric combinations of subject-based evaluation, including the top F1 score on every one of the five datasets. Ablations attribute large gains to the difference attention (e.g., accuracy on TDBRAIN falls from 91.04% to 70.52% when it is removed) and to the multi-resolution graph learning. If the results hold, MedGNN offers a general high-accuracy backbone for cardiac and neurological signal classification, with a precision-recall balance suited to clinical screening.

What carries the argument

The load-bearing mechanism is the multi-resolution adaptive graph, built by running 1D convolutions with multiple kernel sizes over the input channels to produce embeddings at several temporal resolutions, then assigning each resolution its own learnable adjacency matrix over the channel nodes. Three companion modules carry the rest of the argument: Difference Attention Networks apply self-attention to the first-order finite difference along the time axis (with a residual connection to the original embedding), which suppresses slow baseline drifts; Frequency Convolution Networks apply a learnable element-wise multiplication in the Fourier domain and transform back, capturing frequency information lost in difference space; and the Multi-resolution Graph Transformer applies local attention and graph convolution within each resolution, then average-pools across resolutions before a linear classifier. The learned adjacency matrices are the paper's evidence that channel correlations genuinely differ by scale.

What would settle it

Re-run all ten baselines on the same five datasets behind the reported train-validation-test subject splits, giving each baseline a per-dataset hyperparameter search with the same budget and grid size reported for MedGNN in Appendix A.4, and test for statistical significance across multiple seeds. If a baseline such as Medformer or iTransformer then matches or exceeds MedGNN on the TDBRAIN or PTB-XL benchmarks, the central claim of architectural superiority fails.

Watch

Extended reading notes

Core claim

The central claim is that a model built around the structure of medical signals can beat generic existing time series architectures on ECG and EEG classification. The paper's discovery is that three inductive biases matter together: multi-resolution channel graphs (each temporal scale gets its own learned adjacency matrix over channels), difference attention (multi-head self-attention on the finite difference along time, with the original embedding added back, which cancels baseline wander), and frequency convolution (Fourier transform, learnable complex kernel, inverse transform) for a complementary view. The paper reports the strongest empirical support on five real-world datasets: in sample-based evaluation on ADFD, MedGNN reaches 98.42% accuracy against 97.62% for the closest baseline Medformer; in subject-based evaluation it claims 22 top-1 and 2 top-2 results out of 25, and the leading F1 score on all five datasets. The authors further show that removing difference attention sharply degrades accuracy (from 91.04% to 70.52% on TDBRAIN) and that the learned adjacency matrices differ across resolutions, indicating scale-dependent channel relationships.

Load-bearing premise

The comparison is fair: the ten baselines were run with tuning effort and data splits comparable to MedGNN's, so the reported margins come from the architecture rather than from unequal hyperparameter selection.

Editorial extensions

If this is right

  • MedGNN provides a reusable architecture for EEG and ECG classification tasks beyond the five datasets it tests, potentially aiding early detection of arrhythmia, myocardial infarction, Alzheimer's disease, and Parkinson's disease.
  • Attention on temporal differences offers a built-in robustness to baseline drift, reducing reliance on explicit preprocessing or denoising steps for such signals.
  • The per-resolution adjacency matrices give interpretable channel-connectivity estimates that shift across scales, which could help clinicians see which brain or heart regions interact at which time scales.
  • Top F1 scores on every dataset indicate the model balances precision and recall, which matters in screening settings where both false alarms and missed cases are costly.
  • The reported moderate memory footprint and training time, together with top accuracy, suggest the model is practical for routine GPU-based deployment.

Reading between the lines

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

  • A natural extension is to view difference attention as an instance of a general principle: applying learned operators to finite differences of a signal, rather than the raw signal, is a cheap way to achieve invariance to slowly varying offsets, which could transfer to wearable-sensor and industrial monitoring data with similar drift artifacts.
  • The scale-dependent adjacency matrices suggest a testable clinical hypothesis not pursued in the paper: differences in per-resolution channel graphs between healthy and diseased subjects could serve as connectivity biomarkers for the underlying conditions.
  • Because the paper does not run paired significance tests, a fair next step is to re-benchmark the baselines with matched per-dataset tuning and report confidence intervals; some of the reported top-1 margins may shrink under that protocol, though the large margins in the ablation studies suggest the core mechanisms would survive.
  • The model's success under the harder subject-based (leave-patients-out) protocol is encouraging for translational use, but independent validation on additional multi-center EEG and ECG databases would test whether the gains generalize beyond the five chosen benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes MedGNN, a multi-resolution spatiotemporal graph learning framework for medical time series classification. The method constructs multi-resolution adaptive graphs from multi-scale 1-D convolutions, applies Difference Attention Networks to finite-difference representations to counter baseline wander, uses Frequency Convolution Networks for frequency-domain information, and fuses these with a Multi-resolution Graph Transformer before classification. Experiments are reported on five medical datasets (three EEG, two ECG) under sample-based and subject-based splits, with ten baselines, ablations, efficiency comparisons, and visualization of learned graphs. The paper claims that MedGNN outperforms all baselines in the sample-based evaluation and achieves 22 top-1 and 2 top-2 results out of 25 in subject-based evaluation.

Significance. If the central empirical claim is supported, MedGNN would be a useful and fairly comprehensive architecture for EEG/ECG classification: it jointly models channel-level spatial dependencies, multi-scale temporal patterns, baseline wander, and frequency-domain information. The paper has several strengths: evaluation on five datasets, a subject-based split that tests generalization to unseen patients, standard deviations over five runs, ablations for each major component, and a public code link. The main weakness is that the quantitative comparison is not yet fully verifiable: the baseline tuning protocol and classification adaptation are unspecified, there are no significance tests, and several reported margins are small with overlapping standard deviations. There are also a few dimensional inconsistencies in the architecture description that need clarification before the method can be reproduced from the text.

major comments (3)
  1. [Sections 5.1.2, A.2, A.4; Tables 1 and 2] The baseline comparison is not yet demonstrably fair. Appendix A.4 gives MedGNN's hyperparameter search ranges, but no analogous tuning protocol, validation-based selection rule, or final per-dataset configuration is given for the ten baselines. In addition, six of the baselines are originally forecasting models, and the paper does not state how they are adapted to classification (e.g., pooling strategy, classification head, output layer). This matters because some margins are small: PTB-XL accuracy is 73.87 for MedGNN versus 73.23 for Crossformer, and APAVA accuracy is 82.60 versus 82.35 for SimTSC, with overlapping standard deviations. To support the central claim that MedGNN outperforms all ten baselines, the authors should report per-dataset baseline hyperparameters and tuning budgets, describe the classification adaptation, and provide paired significance tests or confidence intervals for the reported margins.
  2. [Section 4.2, Eqs. (4)-(7)] There is a dimensional inconsistency in the residual connection. Equation (4) defines X'_i(m) as having shape C x (T_m+1) after padding and produces D_i(m) with T_m temporal positions; the difference attention output X_DSA in Eq. (6) therefore has T_m positions, but Eq. (7) adds X'_i(m), which has T_m+1. Please specify whether the residual operates on the padded representation and how the final T_m positions are recovered, or define the difference operator so that it preserves shape.
  3. [Section 4.4, Eq. (15)] The average pooling across resolutions is not well defined because each X_GT^{(m)} has shape C x T_m and T_m = floor(T/k_m) differs with kernel size. The text does not state how tensors of different temporal lengths are aligned, padded, or truncated before pooling. Since this is the fusion step that produces the final representation, the exact alignment procedure must be specified.
minor comments (4)
  1. [Section 5.2.2] The claim of '22 top-1 and 2 top-2 out of 25' appears to treat the tied PTB-XL recall value (61.13 for both MedGNN and Crossformer) as top-2 rather than top-1; please state the tie-breaking convention explicitly.
  2. [Section 5.3, Table 3] The ablation table reports only single values without standard deviations or significance tests, although Tables 1 and 2 report five-run standard deviations; adding error bars or a paired test would make the claimed improvements more convincing.
  3. [Abstract and Section 1] The code link is stated as available, but a versioned archive or commit hash would be needed for reproducibility; please also note the typo 'machanisms' in Section 1 and 'perfomance' in Section 5.2.2.
  4. [Figure 1 caption and Figure 4] There are minor typos: 'and and' appears in the Figure 1 caption, and 'Medforemr' appears in Figure 4; these should be corrected.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: MedGNN's performance claim is measured on held-out subjects against external baselines, not derived from its own fitted assumptions.

full rationale

MedGNN is an empirical architecture paper. Its central claim (Section 5.2, Tables 1 and 2) is that MedGNN outperforms ten baselines on held-out test sets; the subject-based splits are inherited from Medformer [39], an external prior work with no author overlap, and each reported number is an external benchmark outcome rather than a quantity reconstructed from a fitted parameter. The method sections (Section 4.1-4.4) define the multi-resolution graph construction, difference attention, frequency convolution, and graph transformer with explicit equations (Eq. 2-16); none of these equations defines its output in terms of the classification labels or the reported accuracy, so there is no self-definitional reduction. The ablations (Section 5.3, Table 3, Figures 2-3) compare full versus ablated versions on the same held-out evaluation, which is an internal consistency check, not a circular prediction. The paper does cite the authors' own prior frequency-domain and graph work ([9, 10, 11, 44, 45, 46, 47]), and FourierGNN [46] is both a self-cited method and a baseline, but the MedGNN frequency convolution formula (Eq. 8-10) is fully specified in the paper and its stated motivation (multi-view temporal representations) does not rest on those citations as evidence. The skeptical concern that baselines may not have received equal hyperparameter tuning or explicit classification-head adaptation is a correctness/fairness threat to the empirical comparison, not a circularity: even if the comparison were unfair, the reported outcome would still not be equivalent to the model's inputs by construction. A separate internal inconsistency (Eq. 15 average-pools representations with differing T_m without specifying padding) is an implementation defect, not a circular step. No load-bearing step reduces a claimed result to its own inputs, so the circularity score is low.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

MedGNN carries no closed-form derivation; its claim is empirical. The cost is carried by a stack of modeling choices: adaptive multi-resolution graph structure, finite-difference attention, frequency-domain convolution, and per-dataset hyperparameter selection. None of these modules is independently verified against physiological ground truth, so the accuracy gains in Tables 1 and 2 are the only evidence that the machinery works.

free parameters (6)
  • number of resolutions M = subset of {2,4,6,8,10,12,14,16}
    Per-dataset validation choice; the exact value and the actual kernel sizes for each dataset are not reported, so the multi-resolution configuration is not reproducible from text alone.
  • graph node dimension = one of {6,8,10}
    Selected per dataset (Appendix A.4); controls the capacity of channel embeddings and affects all graph transformer layers.
  • encoder layers L = one of {4,6}
    Selected per dataset; defines the depth of the multi-resolution graph transformer.
  • embedding dimension D = one of {256,512}
    Base model dimension, selected per dataset but not reported per dataset.
  • kernel sizes K_1,...,K_M = not reported
    The multi-scale construction depends on M kernel sizes, but only a candidate set of resolutions is mentioned, not the actual kernel values used in final models.
  • batch size = one of {32,64,128,256}
    Optimization hyperparameter selected per dataset.
assumptions (5)
  • domain assumption First-order finite differences remove or reduce baseline wander while preserving classification-relevant temporal patterns.
    Section 4.2 builds Difference Attention entirely on this premise; only indirect ablation support is given, with no direct comparison against standard high-pass filtering or baseline-wander-annotated data.
  • domain assumption Adaptive fully-connected graphs at multiple temporal resolutions are an adequate representation of channel-level spatial dependencies.
    Section 4.1 initializes every resolution graph as fully connected and learns adjacency; the paper does not validate learned edges against known functional connectivity.
  • domain assumption Randomly shuffling samples from the same subject into train and test sets is a reasonable real-world medical setting despite information leakage.
    Section 5.2.1 acknowledges the leakage problem yet still uses sample-based ADFD results (Table 1) as a demonstration of superiority.
  • domain assumption The subject-based split protocol and data preprocessing inherited from Medformer are appropriate and comparable across all baselines.
    Sections 5.1.1 and A.4 rely on this equivalence, but the actual data splits and preprocessing code are not released.
  • standard math The continuous Fourier integral notation in Eq. (8) is an acceptable representation of the discrete Fourier transform applied to finite signals.
    The notation is mathematically imprecise for discrete sequences; a discrete Fourier transform is presumably intended but never stated.
invented entities (2)
  • Multi-resolution adaptive adjacency matrices A_i(m)
    purpose: Model dynamic channel-level dependencies at each temporal resolution.
    Introduced in Section 4.1.2; the learned graphs have no external falsifiable handle and the visualizations in Figures 5 and 6 are not compared against known physiological connectivity.
  • Difference attention representation X_DA
    purpose: Temporal modeling robust to baseline wander.
    Defined in Section 4.2; evaluated only through ablation on downstream accuracy, not through direct measurement of baseline-wander suppression.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MedGNN: Towards Multi-resolution Spatiotemporal Graph Learning for Medical Time Series Classification." pith.science (2026). https://pith.science/paper/CFQBNRY3

@misc{pith2026250204515,
  author       = {Pith},
  title        = {Pith review of: MedGNN: Towards Multi-resolution Spatiotemporal Graph Learning for Medical Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CFQBNRY3}},
  note         = {Machine review of arXiv:2502.04515}
}
read the original abstract

Medical time series has been playing a vital role in real-world healthcare systems as valuable information in monitoring health conditions of patients. Accurate classification for medical time series, e.g., Electrocardiography (ECG) signals, can help for early detection and diagnosis. Traditional methods towards medical time series classification rely on handcrafted feature extraction and statistical methods; with the recent advancement of artificial intelligence, the machine learning and deep learning methods have become more popular. However, existing methods often fail to fully model the complex spatial dynamics under different scales, which ignore the dynamic multi-resolution spatial and temporal joint inter-dependencies. Moreover, they are less likely to consider the special baseline wander problem as well as the multi-view characteristics of medical time series, which largely hinders their prediction performance. To address these limitations, we propose a Multi-resolution Spatiotemporal Graph Learning framework, MedGNN, for medical time series classification. Specifically, we first propose to construct multi-resolution adaptive graph structures to learn dynamic multi-scale embeddings. Then, to address the baseline wander problem, we propose Difference Attention Networks to operate self-attention mechanisms on the finite difference for temporal modeling. Moreover, to learn the multi-view characteristics, we utilize the Frequency Convolution Networks to capture complementary information of medical time series from the frequency domain. In addition, we introduce the Multi-resolution Graph Transformer architecture to model the dynamic dependencies and fuse the information from different resolutions. Finally, we have conducted extensive experiments on multiple medical real-world datasets that demonstrate the superior performance of our method. Our Code is available.

Figures

Figures reproduced from arXiv: 2502.04515 by the authors.

Figure 1
Figure 1. The overall architecture of MedGNN. (a) Multi-resolution graph construction is utilized to learn the dynamic [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Ablation study of Frequency Convolution Networks [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Model effectiveness and efficiency comparison on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Adjacent matrices of multi-resolution graphs learned from APAVA dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Adjacent matrices of multi-resolution graphs learned from TDBRAIN dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Effect of different graph node dimension. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 39 canonical work pages

  1. [1]

    Wasyihun Sema Admass and Girmaw Andualem Bogale. 2024. Arrhythmia clas- sification using ECG signal: A meta-heuristic improvement of optimal weighted feature integration and attention-based hybrid deep learning model. Biomed. Signal Process. Control. 87, Part B (2024), 105565

    work page 2024

  2. [2]

    Bagnall, Jason Lines, Aaron Bostrom, James Large, and Eamonn J

    Anthony J. Bagnall, Jason Lines, Aaron Bostrom, James Large, and Eamonn J. Keogh. 2017. The great time series classification bake off: a review and experi- mental evaluation of recent algorithmic advances. Data Min. Knowl. Discov. 31, 3 (2017), 606–660

    work page 2017

  3. [3]

    Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. 2020. Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting. In NeurIPS

    work page 2020

  4. [4]

    Selcan Kaplan Berkaya, Alper Kursat Uysal, Efnan Sora Gunal, Semih Ergin, Serkan Gunal, and M Bilginer Gulmezoglu. 2018. A survey on ECG analysis. Biomedical Signal Processing and Control 43 (2018), 216–235

    work page 2018

  5. [5]

    Manuel Blanco-Velasco, Binwei Weng, and Kenneth E Barner. 2008. ECG signal denoising and baseline wander correction based on the empirical mode decom- position. Computers in biology and medicine 38, 1 (2008), 1–13

    work page 2008

  6. [6]

    Aaron Bostrom and Anthony J. Bagnall. 2017. Binary Shapelet Transform for Multiclass Time Series Classification. Trans. Large Scale Data Knowl. Centered Syst. 32 (2017), 24–46

    work page 2017

  7. [7]

    Michael X Cohen. 2017. Where does EEG come from and what does it mean? Trends in neurosciences 40, 4 (2017), 208–218

    work page 2017

  8. [8]

    J Escudero, Daniel Abásolo, Roberto Hornero, Pedro Espino, and Miguel López

    work page

Show all 56 references
  1. [9]

    Wei Fan, Pengyang Wang, Dongkun Wang, Dongjie Wang, Yuanchun Zhou, and Yanjie Fu. 2023. Dish-TS: A General Paradigm for Alleviating Distribution Shift in Time Series Forecasting. In AAAI. AAAI Press, 7522–7529

    2023

  2. [10]

    Wei Fan, Kun Yi, Hangting Ye, Zhiyuan Ning, Qi Zhang, and Ning An. 2024. Deep frequency derivative learning for non-stationary time series forecasting. arXiv preprint arXiv:2407.00502 (2024)

    2024 arXiv

  3. [11]

    Wei Fan, Shun Zheng, Xiaohan Yi, Wei Cao, Yanjie Fu, Jiang Bian, and Tie-Yan Liu. 2022. DEPTS: Deep Expansion Learning for Periodic Time Series Forecasting. In ICLR. OpenReview.net

    2022

  4. [12]

    Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. 2019. Deep learning for time series classification: a review. Data Min. Knowl. Discov. 33, 4 (2019), 917–963

    2019

  5. [13]

    Qihe Huang, Lei Shen, Ruixin Zhang, Shouhong Ding, Binwu Wang, Zhengyang Zhou, and Yang Wang. 2023. Crossgnn: Confronting noisy multivariate time series via cross interaction refinement. Advances in Neural Information Processing Systems 36 (2023), 46885–46902

    2023

  6. [14]

    Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. 2019. Deep learning for time series classification: a review. Data mining and knowledge discovery 33, 4 (2019), 917–963

    2019

  7. [15]

    Rajendra Acharya

    Mahboobeh Jafari, Afshin Shoeibi, Marjane Khodatars, Sara Bagherzadeh, Ahmad Shalbaf, David López-García, Juan Manuel Górriz, and U. Rajendra Acharya. 2023. Emotion recognition in EEG signals using deep learning methods: A review. Comput. Biol. Medicine 165 (2023), 107450

    2023

  8. [16]

    Webb, Irwin King, and Shirui Pan

    Ming Jin, Huan Yee Koh, Qingsong Wen, Daniele Zambon, Cesare Alippi, Ge- offrey I. Webb, Irwin King, and Shirui Pan. 2023. A Survey on Graph Neural Networks for Time Series: Forecasting, Classification, Imputation, and Anomaly Detection. CoRR abs/2307.03759 (2023)

    2023 arXiv

  9. [17]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  10. [18]

    Dani Kiyasseh, Tingting Zhu, and David A. Clifton. 2021. CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients. In ICML (Proceed- ings of Machine Learning Research, Vol. 139) . PMLR, 5606–5615

    2021

  11. [19]

    Lawhern, Amelia J

    Vernon J. Lawhern, Amelia J. Solon, Nicholas R. Waytowich, Stephen M. Gordon, Chou P. Hung, and Brent J. Lance. 2016. EEGNet: A Compact Convolutional Network for EEG-based Brain-Computer Interfaces. CoRR abs/1611.08024 (2016)

    2016 arXiv

  12. [20]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. In ICLR (Poster)

    2018

  13. [21]

    Jason Lines and Anthony J. Bagnall. 2015. Time series classification with en- sembles of elastic distance measures. Data Min. Knowl. Discov. 29, 3 (2015), 565–592

    2015

  14. [22]

    Huaiyuan Liu, Donghua Yang, Xianzhang Liu, Xinglei Chen, Zhiyu Liang, Hongzhi Wang, Yong Cui, and Jun Gu. 2024. Todynet: temporal dynamic graph neural network for multivariate time series classification. Information Sciences (2024), 120914

    2024

  15. [23]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. itransformer: Inverted transformers are effective for time series forecasting. International Conference on Learning Representations (2024)

    2024

  16. [24]

    Andreas Miltiadous, Katerina D Tzimourta, Theodora Afrantou, Panagiotis Ioan- nidis, Nikolaos Grigoriadis, Dimitrios G Tsalikakis, Pantelis Angelidis, Markos G Tsipouras, Euripidis Glavas, Nikolaos Giannakeas, et al. 2023. A dataset of scalp EEG recordings of Alzheimer’s disea...

    2023

  17. [25]

    Mohammad Amin Morid, Olivia R Liu Sheng, and Joseph Dunbar. 2023. Time series prediction using deep learning methods in healthcare. ACM Transactions on Management Information Systems 14, 1 (2023), 1–29

    2023

  18. [26]

    Alharbi, and Yousef Ajami Alotaibi

    Valentin Moy, Mohammed Sidi Yakoub, Sid-Ahmed Selouani, Yasser F. Alharbi, and Yousef Ajami Alotaibi. 2021. Classification of the Neural Correlates of Mind Wandering States in EEG Signals. In TSP. IEEE, 126–129

    2021

  19. [27]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A time series is worth 64 words: Long-term forecasting with transformers. ICLR (2023)

    2023

  20. [28]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing system...

    2019

  21. [29]

    PhysioToolkit PhysioBank. 2000. Physionet: components of a new research resource for complex physiologic signals. Circulation 101, 23 (2000), e215–e220

    2000

  22. [30]

    Andrea L Schaffer, Timothy A Dobbins, and Sallie-Anne Pearson. 2021. In- terrupted time series analysis using autoregressive integrated moving average (ARIMA) models: a guide for evaluating large-scale health interventions. BMC medical research methodology 21 (2021), 1–12

    2021

  23. [31]

    Ahsan Shehzad, Feng Xia, Shagufta Abid, Ciyuan Peng, Shuo Yu, Dongyu Zhang, and Karin Verspoor. 2024. Graph Transformers: A Survey. CoRR abs/2407.09777 (2024)

    2024

  24. [32]

    Michal Teplan et al. 2002. Fundamentals of EEG measurement. Measurement science review 2, 2 (2002), 1–11

    2002

  25. [33]

    Simon L Turner, Amalia Karahalios, Andrew B Forbes, Monica Taljaard, Jeremy M Grimshaw, Allen C Cheng, Lisa Bero, and Joanne E McKenzie. 2020. Design characteristics and statistical methods used in interrupted time series studies evaluating public health interventions: a revie...

    2020

  26. [34]

    Hanneke van Dijk, Guido van Wingen, Damiaan Denys, Sebastian Olbrich, Ros- alinde van Ruth, and Martijn Arns. 2022. The two decades brainclinics research archive for insights in neurophysiology (TDBRAIN) database. Scientific data 9, 1 (2022), 333

    2022

  27. [35]

    Thomas Vincent, Laurent Risser, and Philippe Ciuciu. 2010. Spatially adaptive mixture modeling for analysis of fMRI time series. IEEE transactions on medical imaging 29, 4 (2010), 1059–1074

    2010

  28. [36]

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. 2020. PTB-XL, a large publicly available electrocardiography dataset. Scientific data 7, 1 (2020), 1–15

    2020

  29. [37]

    Yihe Wang, Yu Han, Haishuai Wang, and Xiang Zhang. 2023. Contrast Everything: A Hierarchical Contrastive Framework for Medical Time-Series. In NeurIPS

    2023

  30. [39]

    Yihe Wang, Nan Huang, Taida Li, Yujun Yan, and Xiang Zhang. 2024. Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification. arXiv preprint arXiv:2405.19363 (2024)

    2024 arXiv

  31. [40]

    Zekai Wang, Stavros Stavrakis, and Bing Yao. 2023. Hierarchical deep learning with Generative Adversarial Network for automatic cardiac diagnosis from ECG signals. Comput. Biol. Medicine 155 (2023), 106641

    2023

  32. [41]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. Advances in Neural Information Processing Systems 34 (2021), 22419–22430

    2021

  33. [42]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. 2020. Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks. In KDD. 753–763

    2020

  34. [43]

    Brandon Westover, and Jimeng Sun

    Chaoqi Yang, M. Brandon Westover, and Jimeng Sun. 2023. ManyDG: Many- domain Generalization for Healthcare Applications. In ICLR. OpenReview.net

    2023

  35. [44]

    Kun Yi, Jingru Fei, Qi Zhang, Hui He, Shufeng Hao, Defu Lian, and Wei Fan

  36. [45]

    Kun Yi, Qi Zhang, Longbing Cao, Shoujin Wang, Guodong Long, Liang Hu, Hui He, Zhendong Niu, Wei Fan, and Hui Xiong. 2023. A Survey on Deep Learning based Time Series Analysis with Frequency Transformation.CoRR abs/2302.02173 (2023)

    2023 arXiv

  37. [46]

    Kun Yi, Qi Zhang, Wei Fan, Hui He, Liang Hu, Pengyang Wang, Ning An, Long- bing Cao, and Zhendong Niu. 2024. FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective. Advances in Neural Information Processing Systems 36 (2024)

    2024

  38. [47]

    Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, and Zhendong Niu. 2024. Frequency-domain MLPs are more effective learners in time series forecasting. Advances in Neural Information Processing Systems 36 (2024)

    2024

  39. [48]

    Raneen Younis, Abdul Hakmeh, and Zahra Ahmadi. 2024. MTS2Graph: Inter- pretable multivariate time series classification with temporal evolving graphs. Pattern Recognit. 152 (2024), 110486

    2024

  40. [49]

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2018. Spatio-Temporal Graph Convo- lutional Networks: A Deep Learning Framework for Traffic Forecasting. In IJCAI. 3634–3640

    2018

  41. [50]

    Daochen Zha, Kwei-Herng Lai, Kaixiong Zhou, and Xia Hu. 2022. Towards Similarity-Aware Time-Series Classification. In SDM. SIAM, 199–207

    2022

  42. [51]

    Daochen Zha, Kwei-Herng Lai, Kaixiong Zhou, and Xia Hu. 2022. Towards similarity-aware time-series classification. In Proceedings of the 2022 SIAM Inter- national Conference on Data Mining (SDM) . SIAM, 199–207

    2022

  43. [52]

    Xiang Zhang, Marko Zeman, Theodoros Tsiligkaridis, and Marinka Zitnik. 2022. Graph-Guided Network for Irregularly Sampled Multivariate Time Series. In ICLR. OpenReview.net

    2022

  44. [53]

    Yunhao Zhang and Junchi Yan. 2022. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. In The Eleventh International Conference on Learning Representations

    2022

  45. [54]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin

  46. [2006]

    Physiological measurement 27, 11 (2006), 1091

    Analysis of electroencephalograms in Alzheimer’s disease patients with multiscale entropy. Physiological measurement 27, 11 (2006), 1091

    2006

  47. [2022]

    In International Conference on Machine Learning

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International Conference on Machine Learning. PMLR, 27268– 27286. A Experimental Details A.1 Datasets The ADFD [24] dataset includes 88 subjects and contains 69,762 multivariate EEG sampl...

    2025

  48. [2024]

    arXiv preprint arXiv:2411.01623 (2024)

    FilterNet: Harnessing Frequency Filters for Time Series Forecasting. arXiv preprint arXiv:2411.01623 (2024). WWW ’25, April 28-May 2, 2025, Sydney, NSW, Australia Wei Fan, Jingru Fei, Dingyu Guo, Kun Yi, Xiaozhuang Song, Haolong Xiang, Hangting Ye, and Min Li

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.