Pith. sign in

REVIEW 3 major objections 6 minor 30 references

Dual-Task Graph Neural Network for Joint Seizure Onset Zone Localization and Outcome Prediction using Stereo EEG

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A dual-task graph neural network jointly predicts post-surgical seizure freedom and localizes the seizure onset zone from stereo EEG windows.

desk verdict The dual-task framing is novel and the pediatric sEEG dataset is a real asset, but the headline accuracies are inflated by window-level leakage and the honest LOOCV numbers are far weaker. read the letter →

arxiv 2505.23669 v1 pith:WDBUXDD7 submitted 2025-05-29 eess.SP q-bio.NC

Syeda Abeera Amir , Artur Agaronyan , William Gaillard , Chima Oluigbo , Syed Muhammad Anwar This is my paper
classification eess.SPq-bio.NC
keywords seizureonsetzonedrug-resistantepilepsystereoEEGgraphneuralnetworkdual-tasklearningfreedompredictionfunctionalconnectivitypediatric
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 claims that a single graph neural network can perform two clinically intertwined jobs on stereo EEG (sEEG) recordings from children with drug-resistant epilepsy: predict whether ablative surgery will make the patient seizure-free, and identify which electrode contacts lie in the seizure onset zone (SOZ). The authors argue the two tasks should be trained jointly rather than separately, because precise localization is presumed to be what drives good surgical outcomes. Each non-overlapping 10-second window becomes a connectivity graph, and window-level 10-fold cross-validation gives a reported 89.31% accuracy on seizure-freedom prediction and 94.72% accuracy on SOZ channel localization. If these numbers hold, one pipeline could present a surgeon with both a patient-level prognosis and a channel-level target map.

What carries the argument

The load-bearing object is the dual-task graph convolutional network that operates on per-window functional connectivity graphs. It computes node embeddings $\mathbf{H} \in \mathbb{R}^{N \times D}$ through message passing; the node head classifies every channel as SOZ or non-SOZ, while the graph head pools the embeddings, concatenates them with the six global graph descriptors, and emits one seizure-freedom prediction per window. The two heads are coupled by the weighted cross-entropy loss $L_{\text{total}} = (1-\alpha) L_{\text{graph}} + \alpha L_{\text{node}}$, whose balance parameter $\alpha$ is tuned by hyperparameter search; this shared loss is the mechanism by which the paper claims the two tasks help each other. The coupling claim rests on the ablation in which omitting the global descriptors lowers node-level accuracy from 0.94 to 0.86.

What would settle it

Run the evaluation with strictly patient-disjoint folds: train on all windows from a subset of the 20 patients and test only on windows from patients never seen in training, reporting the abstract's window-level metrics. The paper already offers one such comparison in its patient-wise leave-one-out column, where graph-level accuracy drops from 0.89 to 0.87 and node-level from 0.94 to 0.88 with standard deviations near 0.2, so a faithful patient-disjoint protocol would show whether the headline accuracies survive cross-patient generalization.

Watch

Extended reading notes

Core claim

The central claim is that optimizing a node-level SOZ classification head and a graph-level outcome head together in one GNN, with a shared weighted loss, yields accurate joint predictions from per-window functional connectivity graphs. Each window's graph has channels as nodes, thresholded Pearson correlations ($\tau = 0.3$) as edges, 142-dimensional per-node features (spectral power, statistical moments, Hjorth parameters, wavelet energies, local graph metrics), and six global topology descriptors (global efficiency, characteristic path length, average clustering, modularity, small-world index, assortativity). Two graph-convolution layers produce node embeddings; the node head classifies each channel as SOZ or not, while the graph head mean-pools the embeddings, appends the global descriptors, and classifies the window's outcome. Under window-level 10-fold cross-validation the paper reports 89.31% graph-level accuracy for seizure-freedom and 94.72% node-level accuracy for SOZ localization. The ablations support a specific mechanistic claim: removing the six global graph descriptors drops node-level accuracy from 0.94 to 0.86, which the authors take as evidence that whole-network statistics feed back into channel-level localization.

Load-bearing premise

The load-bearing premise is that non-overlapping 10-second windows cut from the same patient's recording behave as independent samples, so that window-level cross-validation measures generalization to new data rather than memorization of each patient's signal patterns.

Editorial extensions

If this is right

  • A single inference pass over an sEEG window would deliver both a channel-level SOZ map and a patient-level prognosis, compacting what currently takes separate pipelines into one pre-surgical readout.
  • The ablation result implies that network-wide connectivity statistics are necessary for accurate channel-level localization, not just for outcome prediction, so future models should keep global and local features in a shared pathway.
  • The network analysis yields a testable prognostic pattern: patients without seizure freedom show denser, more clustered, more synchronized SOZ networks at onset (SOZ degree 5.4 vs 2.25, clustering 0.708 vs 0.256), suggesting onset-connectivity strength may itself be outcome-relevant.
  • Because the current edges capture only zero-lag linear correlation, the paper's stated next step is sequence-aware or directed connectivity modeling, which could capture propagation delays that static correlation graphs miss.

Reading between the lines

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

  • If the window-independence premise fails, the headline window-level accuracies largely reflect within-patient memorization; the paper's own patient-wise leave-one-out column (graph 0.87 ± 0.252, node 0.88 ± 0.204) already shows the numbers drop and turn noisy when whole patients are held out, and the paper itself concedes the single-center 20-patient cohort limits statistical power.
  • The paper reports no non-graph baselines on the same features; a logistic regression or random forest on the 142-dimensional node features would separate the contribution of the GNN architecture from the contribution of the feature set.
  • A natural extension would train the same dual-task model on pre-ictal or interictal windows instead of ictal ones; if localization accuracy survives, the method would support planning before a seizure occurs, parallel to interictal-based SOZ localization.
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 / 6 minor

Summary. The paper proposes a dual-task graph neural network that, from windowed stereo-EEG (sEEG) recordings of 20 pediatric epilepsy patients, jointly predicts post-surgical seizure-freedom outcome (graph-level) and localizes seizure onset zone channels (node-level). For each 10-second window, node features (spectral, statistical, Hjorth, wavelet, local graph) and six global graph descriptors are extracted, and a thresholded Pearson-correlation adjacency matrix is used as the graph. The model is trained with a weighted combination of graph- and node-level cross-entropy losses, with hyperparameters tuned via Optuna. The authors report high accuracies under 'seizure-wise' 10-fold cross-validation and lower, high-variance accuracies under patient-wise leave-one-out cross-validation, and they include ablation studies and a descriptive network analysis. The central claim is that jointly modeling outcome and SOZ improves accuracy and interpretability, supporting AI-assisted epilepsy surgery planning.

Significance. If the reported performance held for new patients, the dual-task architecture would be a clinically relevant contribution: it would combine two tasks that are usually handled separately, on a pediatric sEEG dataset with deep electrode placements, which is underrepresented in the literature. The paper also includes a patient-wise LOOCV evaluation that is free of the window-level leakage affecting the headline numbers, and the authors candidly acknowledge the small cohort and overfitting in the Limitations section. That said, the significance is conditional: the abstract and conclusions emphasize the leakage-prone seizure-wise numbers, and the patient-wise results, which are the appropriate generalization estimate, are substantially weaker and highly variable (accuracy 0.87 ± 0.252 graph-level, 0.88 ± 0.204 node-level). The study is best viewed as a pilot demonstration needing stronger evaluation before its clinical claims are accepted.

major comments (3)
  1. [Abstract and Section 3, Table 1] The headline accuracies (89.31% graph-level, 94.72% node-level) come from 'seizure-wise' 10-fold cross-validation applied to sEEG windows. Because Section 2.2 segments each patient's continuous recording into non-overlapping 10-second windows that all inherit the same patient-level outcome label and the same per-channel SOZ mask, a random window-level split places highly correlated windows from the same patient and electrode montage into both training and test folds. These numbers therefore reflect within-patient memorization rather than cross-patient generalization. The patient-wise LOOCV in the same table (0.87 ± 0.252 and 0.88 ± 0.204) is the appropriate leakage-free estimate and shows markedly lower, high-variance performance. The Abstract's claim and the Section 4 statement that the model is effective 'in generalizing across patients' are not supported by the seizure-wise protocol; the paper should be re-centered on the patient-wise results.
  2. [Section 3, first paragraph] The evaluation protocol is described inconsistently: the text says '10-fold seizure-level cross-validation, training on 60% of the sEEG segments, while splitting the rest of the 40% for validation and testing.' A 10-fold cross-validation does not use a 60/40 split, and it is unclear how the validation portion was used (e.g., early stopping or hyperparameter selection) and whether folds were defined per patient, per seizure, or per window. The term 'seizure-level' is also used interchangeably with 'segment-level' later in the paper. Please specify the exact splitting scheme, the number of folds, and how the reported mean and standard deviation across folds were computed; otherwise the reported 0.89 ± 0.097 and 0.94 ± 0.004 cannot be reproduced or interpreted.
  3. [Section 4 and Limitations] The authors concede in the Limitations paragraph that 'the model tends to overfits on some patients while underfitting on others' and that segment-level cross-validation does not mitigate this. This concession, together with the large standard deviations in the patient-wise LOOCV results (0.252 and 0.204 in Table 1), directly contradicts the Discussion's assertion that the model is robust across patient-wise analysis and the Conclusion's claim that 'the robustness of our model across both seizure-wise and patient-wise analyses suggests that it can generalize well to diverse patient populations.' The generalization claims should be tempered to match the observed variance, and the paper should report per-patient results (e.g., per-patient accuracies or a confusion matrix) so that readers can judge the extent of the overfitting.
minor comments (6)
  1. [Abstract] The error bars in the Abstract are inconsistent with Table 1: the abstract reports 89.31 ± 0.0976 % and 94.72 ± 0.0041 %, while Table 1 lists 0.89 ± 0.097 and 0.94 ± 0.004. Please use a consistent convention (fractions versus percentage points) and correct the apparent factor-of-100 discrepancy in the graph-level standard deviation.
  2. [Section 2.2] The sentence 'A sampling frequency of 128 Hz was used to downsample the features and remove redundancies' is unclear — it is presumably the sEEG signal, not the features, that is downsampled to 128 Hz before computing the power spectral density. Please clarify.
  3. [Contributions bullet list] There is a typo in the first contribution: 'disgnosed' should be 'diagnosed.'
  4. [Section 3.1 and Table 3] The network analysis is based on two illustrative patients (one per class), and Table 3 reports descriptive metrics for three windows without error bars or statistical testing. The Discussion generalizes from these examples to statements about differences between seizure-free and non-seizure-free connectivity patterns; these should be presented explicitly as illustrative observations, not as confirmatory findings.
  5. [Table 3] For class 0, window W3, the reported average SOZ PLV is 0.000 and average clustering is 0.000, which seems implausible if SOZ nodes are present in the network; please verify the calculation or clarify what this entry means.
  6. [References] Reference [14] is a prior work by the same group on graph-based sEEG analysis for seizure-freedom prediction. The authors should clarify the relationship between the two datasets and the novelty of the present work relative to that paper, including whether the patient cohorts overlap.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation-chain circularity; the reported accuracies are empirical cross-validation results, and the self-citations are not load-bearing.

full rationale

This is an empirical supervised-learning paper, not a derivation from first principles. The central claims (graph-level 89.31% and node-level 94.72% accuracy) are measured under cross-validation rather than deduced from the model definition. The loss function Ltotal = (1-α)Lgraph + αLnode, the feature definitions, and the graph construction are all standard and are not defined in terms of the reported accuracies. No fitted parameter is renamed as a prediction: hyperparameters are selected by Optuna inside the CV loop, and the reported test metrics are computed on held-out windows. The paper cites prior work by the same group ([10], [11], [13], [14]); these citations are contextual background, and the present claims do not rely on any theorem or result from them. A potential concern is that window-level 10-fold CV uses non-overlapping 10s windows from the same patient in both training and test partitions, and each window inherits the patient's fixed outcome label and SOZ mask; this can inflate accuracy through patient-specific memorization, as reflected in the much lower, high-variance patient-wise LOOCV (0.87±0.252 graph, 0.88±0.204 node). That is a statistical validity/leakage issue, not a circularity of the kind where the prediction is equivalent to its input by construction, since the test labels are not entered into the training objective. Accordingly, no circular step is exhibited, and the circularity score is minimal.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The model introduces no new physical or biological entities; it recombines standard features (PSD, Hjorth, wavelet, graph metrics) and a standard GCN. All free parameters are hyperparameters or fixed thresholds, none of which are provided with enough detail to reproduce.

free parameters (5)
  • Task weight alpha = not reported
    Balances the node-level and graph-level loss in L_total = (1-alpha) L_graph + alpha L_node; tuned via Optuna but no chosen value is given.
  • Correlation threshold tau = 0.3
    Threshold for building the sparse adjacency matrix from Pearson correlations; chosen by hand without a reported sensitivity analysis.
  • Window length W = 10 s
    Segmentation window length; no comparison to other window sizes is reported.
  • PLV threshold theta = 0.65
    Used in the network analysis visualization (Section 3.1); fixed value with no stated basis.
  • Optuna hyperparameters = not reported
    The paper mentions 100 trials and 30 epochs per trial, but does not list the final learning rate, hidden dimensions, number of layers, or other architecture choices.
assumptions (3)
  • domain assumption Non-overlapping 10-second windows from a patient are exchangeable samples for cross-validation
    Section 2.2 segments continuous recordings into windows, and Section 3 applies window-level cross-validation. This is the load-bearing assumption behind the reported accuracies.
  • domain assumption Zero-lag Pearson correlation thresholded at tau captures clinically relevant functional connectivity
    Section 2.2 constructs adjacency from thresholded correlations; the paper itself acknowledges this ignores directionality and nonlinearity.
  • domain assumption Clinical SOZ annotations and Engel outcomes are accurate ground truth
    Section 2.1 relies on expert labels and 6-month Engel follow-up; no inter-rater reliability or independent verification is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-Task Graph Neural Network for Joint Seizure Onset Zone Localization and Outcome Prediction using Stereo EEG." pith.science (2026). https://pith.science/paper/WDBUXDD7

@misc{pith2026250523669,
  author       = {Pith},
  title        = {Pith review of: Dual-Task Graph Neural Network for Joint Seizure Onset Zone Localization and Outcome Prediction using Stereo EEG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDBUXDD7}},
  note         = {Machine review of arXiv:2505.23669}
}
abstract

Accurately localizing the brain regions that triggers seizures and predicting whether a patient will be seizure-free after surgery are vital for surgical planning and patient management in drug-resistant epilepsy. Stereo-electroencephalography (sEEG) delivers high-fidelity intracranial recordings that enable clinicians to precisely locate epileptogenic networks. However, the clinical identification is subjective and dependent on the expertise of the clinical team. Data driven approaches in this domain are sparse, despite the fact that sEEG offers high temporal-fidelity related to seizure dynamics that can be leveraged using graph structures ideal for imitating brain networks. In this study, we introduce a dual-task graph-neural network (GNN) framework that operates on windowed sEEG recordings to jointly predict seizure-freedom outcomes and identify seizure-onset-zone (SOZ) channels. We assemble non-overlapping 10 second windows from 51 clinical seizures spread across 20 pediatric patients, with sEEG data annotated by clinical experts. For each temporal window we construct a functional connectivity graph via thresholded Pearson correlations and extract rich node features (spectral, statistical, wavelet, Hjorth and local graph features), alongside six global graph descriptors. We optimize a combined cross-entropy loss with a tunable task-weight, and select model hyper-parameters via Optuna. Under window-level 10-fold cross-validation, the model achieves a mean graph-level accuracy of $89.31 \pm 0.0976 \%$ for seizure-freedom prediction and a node-level SOZ localization accuracy of $94.72. \pm 0.0041 \%$. For the best performing model, we ran additive and leave-one-out ablation studies to explore feature importance for graph and node-level accuracy.

Figures

Figures reproduced from arXiv: 2505.23669 by the authors.

Figure 1
Figure 1. Our proposed Dual-Task Graph Neural Network model architecture. The diagram describes [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Demonstration of (a) sEEG electrode placement within the brain and (b) post-op Computed [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Network dynamics at seizure onset from samples (a) without seizure freedom, (b) with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of the dual-task GNN’s performance for two types of ablation results. (a) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: An example of the SOZ identified by our model representing the posterior insula in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages

  1. [1]

    Analysis of EEG records in an epileptic patient using wavelet transform

    Hojjat Adeli, Ziqin Zhou, and Nahid Dadmehr. Analysis of EEG records in an epileptic patient using wavelet transform. Journal of Neuroscience Methods, 123(1):69–87, February 2003. ISSN 0165-0270. doi: 10.1016/S0165-0270(02)00340-0. URL https://www.sciencedirect. com/science/article/pii/S0165027002003400. Number: 1

    work page doi:10.1016/s0165-0270(02)00340-0 2003

  2. [2]

    Milligan

    Tracey A. Milligan. Epilepsy: A Clinical Overview. The American Journal of Medicine, 134 (7):840–847, July 2021. ISSN 1555-7162. doi: 10.1016/j.amjmed.2021.01.038. Number: 7

    work page doi:10.1016/j.amjmed.2021.01.038 2021

  3. [3]

    Recent advances in epilepsy

    Mark Manford. Recent advances in epilepsy. Journal of Neurology, 264(8):1811–1824, 2017. ISSN 0340-5354. doi: 10.1007/s00415-017-8394-2. URL https://www.ncbi.nlm.nih. gov/pmc/articles/PMC5533817/. Number: 8

    work page doi:10.1007/s00415-017-8394-2 2017

  4. [4]

    Epilepsy and Cognitive Impairment in Childhood and Adolescence: A Mini-Review

    Francesca Felicia Operto, Grazia Maria Giovanna Pastorino, Andrea Viggiano, Giovanni Battista Dell’Isola, Gianluca Dini, Alberto Verrotti, and Giangennaro Coppola. Epilepsy and Cognitive Impairment in Childhood and Adolescence: A Mini-Review. Current Neuropharmacology, 21 (8):1646–1665, 2023. ISSN 1875-6190. doi: 10.2174/1570159X20666220706102708. Number: 8

    work page doi:10.2174/1570159x20666220706102708 2023

  5. [5]

    Miller and Anthony L

    Kai J. Miller and Anthony L. Fine. Decision-making in stereotactic epilepsy surgery. Epilepsia, 63(11):2782–2801, November 2022. ISSN 1528-1167. doi: 10.1111/epi.17381. Number: 11

    work page doi:10.1111/epi.17381 2022

  6. [6]

    Sarat Chandra, Savita Sapra, Sheffali Gulati, Mani Kalaivani, Ajay Garg, Chandra S

    Rekha Dwivedi, Bhargavi Ramanujam, P. Sarat Chandra, Savita Sapra, Sheffali Gulati, Mani Kalaivani, Ajay Garg, Chandra S. Bal, Madhavi Tripathi, Sada N. Dwivedi, Ra- jesh Sagar, Chitra Sarkar, and Manjari Tripathi. Surgery for Drug-Resistant Epilepsy in Children. New England Journal of Medicine , 377(17):1639–1647, October 2017. ISSN 0028-4793. doi: 10.10...

    work page doi:10.1056/nejmoa1615335 2017

  7. [7]

    Interictal SEEG Resting-State Connectivity Localizes the Seizure Onset Zone and Predicts Seizure Outcome

    Haiteng Jiang, Vasileios Kokkinos, Shuai Ye, Alexandra Urban, Anto Bagi´c, Mark Richardson, and Bin He. Interictal SEEG Resting-State Connectivity Localizes the Seizure Onset Zone and Predicts Seizure Outcome. Advanced Science (Weinheim, Baden-Wurttemberg, Germany), 9 (18):e2200887, June 2022. ISSN 2198-3844. doi: 10.1002/advs.202200887. Number: 18

    work page doi:10.1002/advs.202200887 2022

  8. [8]

    The association between structural connectivity and anti-seizure medication response in patients with temporal lobe epilepsy

    Dong Ah Lee, Junghae Ko, Sung-Tae Kim, Ho-Joon Lee, and Kang Min Park. The association between structural connectivity and anti-seizure medication response in patients with temporal lobe epilepsy. Epilepsia Open, 9(6):2408–2418, December 2024. ISSN 2470-9239. doi: 10.1002/epi4.13076. Number: 6

    work page doi:10.1002/epi4.13076 2024

Show all 30 references
  1. [9]

    Niemeyer, Peijuan Luo, Carmen Pons, Shiqiang Wu, Hongtao Ma, Jyun-You Liou, Daniel Surinach, Suhasa B

    James E. Niemeyer, Peijuan Luo, Carmen Pons, Shiqiang Wu, Hongtao Ma, Jyun-You Liou, Daniel Surinach, Suhasa B. Kodandaramaiah, and Theodore H. Schwartz. Seizure network characterization by functional connectivity mapping and manipulation. Neurophotonics, 12 (Suppl 1):S14605, ...

    2025 doi

  2. [10]

    Self-supervised learning for seizure classification using ecog spectrograms

    Van Lam, Chima Oliugbo, Abhijeet Parida, Marius George Linguraru, and Syed Muhammad Anwar. Self-supervised learning for seizure classification using ecog spectrograms. In Medical Imaging 2024: Computer-Aided Diagnosis, volume 12927, pages 606–611. SPIE, 2024

    2024

  3. [11]

    Epileptic seizure classification using multidimensional eeg spectrograms

    John Mutersbaugh, Van Lam, Marius George Linguraur, and Syed Muhammad Anwar. Epileptic seizure classification using multidimensional eeg spectrograms. In 2023 19th International Symposium on Medical Information Processing and Analysis (SIPAIM), pages 1–4. IEEE, 2023

    2023

  4. [12]

    Epileptic seizure prediction using scalp electroencephalogram signals

    Syed Muhammad Usman, Shehzad Khalid, and Zafar Bashir. Epileptic seizure prediction using scalp electroencephalogram signals. Biocybernetics and Biomedical Engineering , 41 (1):211–220, January 2021. ISSN 0208-5216. doi: 10.1016/j.bbe.2021.01.001. URL https: //www.sciencedirec...

    2021 doi

  5. [13]

    Self- supervised graph transformer with contrastive learning for brain connectivity analysis towards improving autism detection

    Yicheng Leng, Syed Muhammad Anwar, Islem Rekik, Sen He, and Eung-Joo Lee. Self- supervised graph transformer with contrastive learning for brain connectivity analysis towards improving autism detection. arXiv preprint arXiv:2501.16346, 2025. 10

    2025 arXiv

  6. [14]

    Graph-based deep learning on stereo eeg for predicting seizure freedom in epilepsy patients

    Artur Agaronyan, Syeda Abeera Amir, Nunthasiri Wittayanakorn, John Schreiber, Marius G Linguraru, William Gaillard, Chima Oluigbo, and Syed Muhammad Anwar. Graph-based deep learning on stereo eeg for predicting seizure freedom in epilepsy patients. arXiv preprint arXiv:2502.15...

    2025 arXiv

  7. [15]

    Graphical Insight: Revolu- tionizing Seizure Detection with EEG Representation

    Muhammad Awais, Samir Brahim Belhaouari, and Khelil Kassoul. Graphical Insight: Revolu- tionizing Seizure Detection with EEG Representation. Biomedicines, 12(6):1283, June 2024. ISSN 2227-9059. doi: 10.3390/biomedicines12061283. Number: 6

    2024 doi

  8. [16]

    A comprehensive overview of graph neural network-based approaches to clustering for spatial transcriptomics

    Teng Liu, Zhao-Yu Fang, Zongbo Zhang, Yongxiang Yu, Min Li, and Ming-Zhu Yin. A comprehensive overview of graph neural network-based approaches to clustering for spatial transcriptomics. Computational and Structural Biotechnology Journal, 23:106–128, December

  9. [17]

    Asuncion, and Appaji Rayi

    Sharanya Ramakrishnan, Ria Monica D. Asuncion, and Appaji Rayi. Localization-Related Epilepsies on EEG. In StatPearls. StatPearls Publishing, Treasure Island (FL), 2025. URL http://www.ncbi.nlm.nih.gov/books/NBK557645/

    2025

  10. [18]

    van ’t Klooster, Sandra M

    Jiaojiao Guo, Ziyi Wang, Maryse A. van ’t Klooster, Sandra M. Van Der Salm, Frans S. Leijten, Kees P. Braun, and Maeike Zijlmans. Seizure Outcome After Intraoperative Electrocorticography-Tailored Epilepsy Surgery: A Systematic Review and Meta-Analysis.Neu- rology, 102(11):e20...

    2024 doi

  11. [19]

    Graph-generative neural network for EEG-based epileptic seizure detection via discovery of dynamic brain func- tional connectivity

    Zhengdao Li, Kai Hwang, Keqin Li, Jie Wu, and Tongkai Ji. Graph-generative neural network for EEG-based epileptic seizure detection via discovery of dynamic brain func- tional connectivity. Scientific Reports, 12:18998, November 2022. ISSN 2045-2322. doi: 10.1038/s41598-022-23...

    2022 doi

  12. [20]

    Synchronization- based graph spatio-temporal attention network for seizure prediction

    Jie Xiang, Yanan Li, Xubin Wu, Yanqing Dong, Xin Wen, and Yan Niu. Synchronization- based graph spatio-temporal attention network for seizure prediction. Scientific Reports, 15: 4080, February 2025. ISSN 2045-2322. doi: 10.1038/s41598-025-88492-5. URL https: //www.ncbi.nlm.nih...

    2025 doi

  13. [21]

    Valiante

    Daniele Grattarola, Lorenzo Livi, Cesare Alippi, Richard Wennberg, and Taufik A. Valiante. Seizure localisation with attention-based graph neural networks. Expert Systems with Applica- tions, 203:117330, October 2022. ISSN 0957-4174. doi: 10.1016/j.eswa.2022.117330. URL https:...

    2022

  14. [22]

    Leveraging interictal multimodal features and graph neural networks for automated planning of epilepsy surgery

    Petr Nejedly, Valentina Hrtonova, Martin Pail, Jan Cimbalnik, Pavel Daniel, V ojtech Travnicek, Irena Dolezalova, Filip Mivalt, Vaclav Kremen, Pavel Jurak, Gregory A Worrell, Birgit Frauscher, Petr Klimes, and Milan Brazdil. Leveraging interictal multimodal features and graph ...

    2025 doi

  15. [23]

    Engel, S

    J. Engel, S. Wiebe, J. French, M. Sperling, P. Williamson, D. Spencer, R. Gumnit, C. Zahn, E. Westbrook, B. Enos, Quality Standards Subcommittee of the American Academy of Neurol- ogy, American Epilepsy Society, and American Association of Neurological Surgeons. Practice param...

    2003

  16. [24]

    Global network properties

    Ralf Steuer and Gorka Zamora López. Global network properties. Analysis of biological networks, pages 29–63, 2008

    2008

  17. [25]

    An overview of power spectral density (psd) calculations

    Richard N Youngworth, Benjamin B Gallagher, and Brian L Stamper. An overview of power spectral density (psd) calculations. Optical manufacturing and testing VI, 5869:206–216, 2005. 11

    2005

  18. [26]

    A novel eeg feature extraction method using hjorth parameter

    Seung-Hyeon Oh, Yu-Ri Lee, Hyoung-Nam Kim, et al. A novel eeg feature extraction method using hjorth parameter. International Journal of Electronics and Electrical Engineering, 2(2): 106–110, 2014

    2014

  19. [27]

    The wavelet transform, volume 4

    Ram Shankar Pathak. The wavelet transform, volume 4. Springer Science & Business Media, 2009

    2009

  20. [28]

    Optuna: A next-generation hyperparameter optimization framework

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623–2631, 2019

    2019

  21. [29]

    J. P. Lachaux, E. Rodriguez, J. Martinerie, and F. J. Varela. Measuring phase synchrony in brain signals. Human Brain Mapping, 8(4):194–208, 1999. ISSN 1065-9471. doi: 10.1002/(sici) 1097-0193(1999)8:4<194::aid-hbm4>3.0.co;2-c. 12

    1999 doi

  22. [2024]

    doi: 10.1016/j.csbj.2023.11.055

    ISSN 2001-0370. doi: 10.1016/j.csbj.2023.11.055

    2001 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.