Pith. sign in

REVIEW 4 major objections 6 minor 45 references

Refined Causal Graph Structure Learning via Curvature for Brain Disease Classification

T0 review · 4 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 The paper claims that brain networks built from transfer entropy and refined by causality-aware curvature rewiring outperform correlation-based and learnable graph classifiers for detecting brain disease, achieving an average F1 of 0.7134…

desk verdict Sensible TE-plus-curvature pipeline, but the reported F1 edge is unverified: one split per dataset, no error bars, and an inconsistent c-selection protocol. read the letter →

arxiv 2506.15708 v1 pith:R7UMNJMN submitted 2025-05-30 cs.LG cs.AI

Falih Gozi Febrinanto , Adonia Simango , Chengpei Xu , Jingjing Zhou , Jiangang Ma , Sonika Tyagi , Feng Xia This is my paper
classification cs.LGcs.AI
keywords braindiseaseclassificationtransferentropycausalgraphrewiringRiccicurvatureneuralnetworksfMRIfunctionalconnectivity
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 representing a brain as a causal graph, built by measuring transfer entropy between fMRI time series of brain regions, carries more disease-relevant information than the correlation-based graphs that dominate graph-neural-network brain classifiers. It introduces CGB (Causal Graphs for Brains), which constructs a directed adjacency matrix from transfer entropy and then rewires the graph with a curvature-based procedure called CSDRF that preserves the strongest causal edges while relieving information bottlenecks. On four public fMRI datasets, COBRE, ACPI, ABIDE, and ADNI, CGB reports an average F1 score of 0.7134, above the best baseline (Bargain) at 0.6801. A sympathetic reader would care because the claim is that biological cause-and-effect information, not merely statistical association, improves automated diagnosis of brain disorders.

What carries the argument

The carrying object is the causal graph adjacency matrix produced by transfer entropy, combined with the balanced Forman curvature operator used in the CSDRF rewiring. Transfer entropy (Eq. 5) quantifies directed information flow from one ROI time series to another and is used both to set edges and to seed node features; the balanced Forman curvature (Eq. 7) measures how much each edge acts as a bottleneck, and CSDRF multiplies the curvature improvement of each candidate edge by the transfer entropy along that candidate direction so that rewiring never introduces causally meaningless connections. The graph convolutional network (Eq. 8) then operates on the refined graph.

What would settle it

Compute transfer entropy on each dataset with the method's unspecified histogram bins and windows, then replace the transfer-entropy weights in the node-feature matrix and in CSDRF's edge-selection score with either Pearson correlations or random permutations of the same weights; if F1 does not drop materially, the causal signal is not the active ingredient. A second check is to train CGB with CSDRF's sampling step driven solely by curvature (the original SDRF rule) and compare F1; if the gain over SDRF vanishes, the causality-information term in the rewiring is doing no work.

Watch

Extended reading notes

Core claim

The paper's central claim is that causal graph structure learned from transfer entropy, refined by a causality-aware discrete Ricci flow, yields brain networks that graph neural networks can classify more accurately than networks built from correlation, thresholding, diffusion, or end-to-end learned graphs. CGB computes transfer entropy $T_{v_j \to v_i}$ between every pair of ROI signals, binarizes it into a directed adjacency matrix via a threshold $c$, and uses the resulting directed graph plus a matrix of pairwise TE values as node features. The CSDRF rewiring step repeatedly finds the edge with the lowest balanced Forman curvature, scores candidate added edges by curvature improvement multiplied by the transfer entropy $T_{v_k \to v_l}$ of the candidate direction, samples one via a softmax over these scores, and removes high-curvature edges whose causal strength is low. The paper reports that this pipeline outperforms eight existing brain-network classifiers on average F1 across the four datasets, and its ablations attribute the gain to the causal graph construction and to the CSDRF refinement.

Load-bearing premise

The graph and node features are built from transfer entropy estimates computed by histogram density estimation, and the paper gives no bin count or embedding-window values and no check that these estimates are stable on fMRI BOLD series of 150 to 700 time points; if the transfer entropy values are noisy or biased, the causal graph and its node features are corrupted and the reported advantage over correlation-based graphs could disappear.

Editorial extensions

If this is right

  • Brain networks built from transfer entropy carry disease-discriminative information beyond correlation; on these four datasets the causal graph beats thresholded correlation graphs and learnable graphs.
  • CSDRF's preservation of causally strong edges during curvature rewiring is what makes the refinement compatible with directed causal graphs; it directly improves on the standard SDRF for this setting.
  • The best control threshold $c$ is small (0.05 to 0.1), meaning only the weakest causal edges should be pruned before refinement.
  • The framework's gain is consistent across four distinct diseases (schizophrenia, ADHD, autism, and Alzheimer's), suggesting transfer-entropy causal connectivity is a general biomarker rather than a dataset-specific artifact.
  • Because node features are also transfer-entropy-derived, the classifier uses causal information twice, once in structure and once in features, making the causal model the entire representation.

Reading between the lines

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

  • If transfer-entropy-based causal graphs generalize, they could be combined with dynamic segmentation of fMRI into snapshots to track how causal flow changes over time and potentially detect earlier disease onset, an extension the authors mention as future work.
  • The same CSDRF principle, multiplying curvature improvement by a domain-relevant edge weight, could be dropped into other non-brain graph learning tasks where edge semantics matter, such as gene regulatory or financial networks.
  • Because transfer entropy captures nonlinear coupling, CGB may be more sensitive to subtle synchronized BOLD dynamics than correlation; one could verify this by comparing against a Granger-causality graph with identical downstream network to isolate the nonlinear contribution.
  • The paper's reliance on supervised labels for choosing the threshold $c$ suggests an unsupervised stability criterion, such as edge persistence across bootstrap resamples, could make the method label-free; the authors note this is beyond scope.
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 CGB, a framework that constructs brain graphs by thresholding transfer entropy estimates between fMRI ROI time series (Eqs. 1-6), then rewires them with a Causality-Informed Stochastic Discrete Ricci Flow (CSDRF) module (Algorithm 1) before training a GCN with multi-head feature transforms and concatenation pooling. The central claim is that CGB outperforms existing brain-disease classification frameworks, with an average F1 of 0.7134 over COBRE, ACPI, ABIDE, and ADNI versus 0.6801 for Bargrain (Tables 2 and 3). The paper also reports ablations, a sensitivity analysis for the threshold c, and qualitative causal-graph visualizations, and it acknowledges limitations including ADNI class imbalance and reliance on supervised labels.

Significance. If the empirical claim were stable, CGB would be a useful engineering contribution: it combines two ideas that are individually plausible (transfer-entropy causal graphs and curvature-based rewiring) and proposes a causality-preserving modification of SDRF. The authors use public datasets, compare with several recent baselines, and include an ablation and a c-sensitivity study. I explicitly credit the transparency about the supervised-learning assumption and the limited interpretability of the heatmaps. However, the manuscript provides no code, no repeated-split statistics, and no specification of several hyperparameters that determine the graphs, so the central performance claim is currently not reproducible from the paper alone. The contribution is therefore significant only conditional on a substantially strengthened experimental protocol.

major comments (4)
  1. [§5.4, Tables 2-3] The central performance claim is supported only by a single 80/20 split per dataset. With test sets of roughly 29 (COBRE), 26 (ACPI), 173 (ABIDE), and 53 (ADNI) samples, the reported F1 margins, especially the 0.033 average-F1 advantage over Bargrain, have no variance estimate. Please report repeated stratified splits or cross-validation with standard deviations, confidence intervals, and fixed random seeds. Without this, the advantage over the best baseline is not statistically verified.
  2. [§5.4 vs §5.7, Eq. (6), Fig. 4] The selection protocol for the threshold c is internally inconsistent and the figure reads as test-set-based selection. Section 5.4 states c=0.1 for COBRE, ABIDE, ADNI and 0.05 for ACPI, while Section 5.7 states that suitable values are 0.1 for ACPI, ABIDE, ADNI and 0.05 for COBRE. Figure 4 plots F1 as a function of c without saying whether these F1 values are computed on validation or test partitions. Because c controls the sparsity of the adjacency matrix in Eq. (6), an overfitted choice could erase part or all of the reported gain. Please specify the exact nested-validation protocol and report the results for the chosen c under that protocol.
  3. [Eqs. (1)-(6), §5.1] The transfer-entropy estimates that determine both the adjacency matrix and the node features are under-specified. Equations (1)-(5) depend on the histogram bin count D and the embedding windows q and o, yet the manuscript only says that q should be 'o or 1' and never reports the values used. Since the BOLD series lengths range from 150 to 700 time points, histogram-based TE estimation is a first-order reliability concern. Please report D, q, and o, and include a stability analysis (e.g., TE values as a function of bin count, or a bootstrap) showing that the resulting graphs are not dominated by estimation noise.
  4. [Algorithm 1, §4.1] Algorithm 1 is not fully specified. It does not state how T_{v_k -> v_l} is obtained for candidate edges (k,l) that are not present in the current graph, nor the values of tau, C+, C-, and the maximum iteration count, nor the convergence criterion. The deletion condition "for edge (i,j) with maximal Ric(i,j) if Ric(i,j) > C+ and T_{v_i -> v_j} < C-" appears to describe removing an edge other than the one just added, and the roles of the two thresholds are not formally defined. A complete specification is needed for reproducibility.
minor comments (6)
  1. [§5.1, Table 1] The ABIDE paragraph says the dataset consists of '11 ROI nodes', while Table 1 lists 111; the text also says 868 patients although 402+464=866. Please correct these numbers.
  2. [Eq. (6)] Equation (6) defines A_ij using T_{v_j -> v_i}, but the surrounding text says the edge records causal influence from v_i to v_j (or vice versa). Please clarify the direction convention so the adjacency matrix and the node features are consistent.
  3. [Algorithm 1] The sets B_1(i) and B_1(j) are not defined in the main text; please state explicitly that they are the 1-hop neighbor sets of nodes i and j.
  4. [Fig. 6] The caption and axis labels of Figure 6 are difficult to parse; please clarify what is plotted (apparently a sliding-window transfer-entropy value over time) and label the source and target nodes consistently with Figure 5.
  5. [Eq. (9)] The multi-head notation in Eq. (9) mixes head projections with head index j in a way that is hard to follow; please define the dimension of each head and show how the concatenation produces h^{(l+1)}_i.
  6. [General] The terms 'geometry curvature' and 'geometric curvature' are used interchangeably; please standardize the terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the CGB pipeline builds graphs and features from transfer entropy and curvature without using classification labels, and the reported gains are not constructed from a fitted constant or a self-citation chain.

full rationale

The central claim is that CGB outperforms baselines in brain disease classification. The derivation chain is: transfer entropy estimates from BOLD signals (Eqs. 1-5) form both the thresholded adjacency matrix (Eq. 6) and the node features; CSDRF rewires this graph using curvature and TE (Algorithm 1); a GCN then produces node embeddings, which are concatenated and passed to a fully connected classifier. Classification labels enter only during supervised training and standard model selection (Section 5.4), not in the graph-construction equations (Eqs. 1-7) or in Algorithm 1. The threshold c is a hyperparameter, not a fitted prediction, and the paper explicitly describes selecting it via validation ('A grid search on validation subsets should be performed', Section 5.4). No equation reduces to another by construction, and no fitted parameter is renamed as a prediction. The self-citations, such as Febrinanto et al. (2023a) for Bargrain and Febrinanto et al. (2023b) for TE-based graphs, are baselines or methodology reuse, not load-bearing uniqueness claims; the TE formula itself is attributed to Schreiber (2000) and curvature rewiring to Topping et al. (2022). The internal inconsistency between Section 5.4 (c = 0.1 for COBRE, 0.05 for ACPI) and Section 5.7 (suitable c = 0.05 for COBRE, 0.1 for ACPI/ABIDE/ADNI) and the unstated TE histogram parameters D, q, and o are reproducibility and correctness risks, not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The graph construction depends on several unvalidated modeling choices: TE thresholds, histogram binning, embedding windows, and CSDRF sampling parameters. These are fitted or left unspecified rather than derived. The core pipeline also assumes TE on BOLD signals reflects meaningful causal interactions, an assumption adopted without empirical support in this paper. No new entities are introduced.

free parameters (6)
  • TE threshold c = 0.1 for COBRE, ABIDE, ADNI; 0.05 for ACPI in Section 5.4 (conflicting with 5.7)
    Controls graph sparsity in Eq. 6; selected by grid search on validation subsets in Section 5.4, making the final graph density a fitted choice.
  • CSDRF temperature tau = not reported
    Controls softmax sampling in Algorithm 1 step 3; not specified in experimental settings, so the rewiring behavior is underspecified.
  • CSDRF thresholds C+ and C- = not reported
    Optional bounds in Algorithm 1 for edge removal; no values are given, so edge-removal behavior is not reproducible.
  • TE histogram bin count D = not reported
    D determines the probability density estimates in Eqs. 1-5; the paper says D is based on the range of values but never states the binning rule used in experiments.
  • TE embedding windows q and o = not reported
    Eq. 5 conditions on q and o past values; the paper mentions Schreiber's suggestion but never states the values used.
  • CSDRF max iterations = not reported
    Algorithm 1 loops until convergence or a maximum number of iterations, but the maximum is not specified.
assumptions (4)
  • domain assumption Transfer entropy, as defined by Eq. 5, is a reliable causal measure for fMRI BOLD time series.
    Invoked throughout Section 4.1; no validation of TE estimation on fMRI is provided.
  • ad hoc to paper The product of curvature gain and transfer entropy in Algorithm 1 is a suitable score for choosing rewiring edges.
    CSDRF rule x_kl = (curvature gain) * T_{vk to vl} is introduced without derivation or ablation against unweighted SDRF.
  • standard math Balanced Forman curvature identifies information bottlenecks that, if rewired, improve GNN expressiveness.
    Adopted from Topping et al. (2022); the paper relies on this theoretical justification in Section 4.1.
  • domain assumption The four public datasets are preprocessed correctly and their parcellations are appropriate for the task.
    Section 5.1 uses the datasets as-is from public repositories; no quality control or preprocessing checks are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Refined Causal Graph Structure Learning via Curvature for Brain Disease Classification." pith.science (2026). https://pith.science/paper/R7UMNJMN

@misc{pith2026250615708,
  author       = {Pith},
  title        = {Pith review of: Refined Causal Graph Structure Learning via Curvature for Brain Disease Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7UMNJMN}},
  note         = {Machine review of arXiv:2506.15708}
}
read the original abstract

Graph neural networks (GNNs) have been developed to model the relationship between regions of interest (ROIs) in brains and have shown significant improvement in detecting brain diseases. However, most of these frameworks do not consider the intrinsic relationship of causality factor between brain ROIs, which is arguably more essential to observe cause and effect interaction between signals rather than typical correlation values. We propose a novel framework called CGB (Causal Graphs for Brains) for brain disease classification/detection, which models refined brain networks based on the causal discovery method, transfer entropy, and geometric curvature strategy. CGB unveils causal relationships between ROIs that bring vital information to enhance brain disease classification performance. Furthermore, CGB also performs a graph rewiring through a geometric curvature strategy to refine the generated causal graph to become more expressive and reduce potential information bottlenecks when GNNs model it. Our extensive experiments show that CGB outperforms state-of-the-art methods in classification tasks on brain disease datasets, as measured by average F1 scores.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [1]

    arXiv preprint arXiv:230815568v6

    Akansha S (2024) Over-squashing in graph neural networks: A comprehensive survey. arXiv preprint arXiv:230815568v6

    work page 2024

  2. [2]

    Machine Learning and Knowledge Extraction 6(1):464--505

    Alsubaie MG, Luo S, Shaukat K (2024) Alzheimer’s disease detection using deep learning on neuroimaging: A systematic review. Machine Learning and Knowledge Extraction 6(1):464--505

    work page 2024

  3. [3]

    In: Learning on Graphs Conference, PMLR, pp 15--1

    Arnaiz-Rodr \' guez A, Begga A, Escolano F, et al (2022) Diffwire: Inductive graph rewiring via the lovász bound. In: Learning on Graphs Conference, PMLR, pp 15--1

    work page 2022

  4. [4]

    Journal of Artificial Intelligence Research 73:767--819

    Assaad CK, Devijver E, Gaussier E (2022) Survey and evaluation of causal discovery methods for time series. Journal of Artificial Intelligence Research 73:767--819

    work page 2022

  5. [5]

    In: International Conference on Machine Learning, PMLR, pp 2528--2547

    Black M, Wan Z, Nayyeri A, et al (2023) Understanding oversquashing in gnns through the lens of effective resistance. In: International Conference on Machine Learning, PMLR, pp 2528--2547

    work page 2023

  6. [6]

    In: Asian Conference on Machine Learning, PMLR, pp 185--200

    Chen W, Wang Y, Du C, et al (2023) Balanced spatial-temporal graph structure learning for multivariate time series forecasting: A trade-off between efficiency and flexibility. In: Asian Conference on Machine Learning, PMLR, pp 185--200

    work page 2023

  7. [7]

    Frontiers in Neuroinformatics 7(27):5

    Craddock C, Benhajali Y, Chu C, et al (2013) The neuro bureau preprocessing initiative: open sharing of preprocessed neuroimaging data and derivatives. Frontiers in Neuroinformatics 7(27):5

    work page 2013

  8. [8]

    Human brain mapping 33(8):1914--1928

    Craddock RC, James GA, Holtzheimer III PE, et al (2012) A whole brain fmri atlas generated via spatially constrained spectral clustering. Human brain mapping 33(8):1914--1928

    work page 2012

Show all 45 references
  1. [9]

    IEEE transactions on medical imaging 42(2):493--506

    Cui H, Dai W, Zhu Y, et al (2022 a ) Braingb: a benchmark for brain network analysis with graph neural networks. IEEE transactions on medical imaging 42(2):493--506

    2022

  2. [10]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, pp 375--385

    Cui H, Dai W, Zhu Y, et al (2022 b ) Interpretable graph neural networks for connectome-based brain disorder analysis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, pp 375--385

    2022

  3. [11]

    Molecular psychiatry 19(6):659--667

    Di Martino A, Yan CG, Li Q, et al (2014) The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism. Molecular psychiatry 19(6):659--667

    2014

  4. [12]

    Tsinghua Science and Technology 28(1):141--149

    Duan Z, Xu H, Huang Y, et al (2022) Multivariate time series forecasting with transfer entropy graph. Tsinghua Science and Technology 28(1):141--149

    2022

  5. [13]

    arXiv preprint arXiv:221108927

    ElGazzar A, Thomas R, Van Wingen G (2022) Benchmarking graph neural networks for fmri analysis. arXiv preprint arXiv:221108927

    2022

  6. [14]

    In: Proceedings of the ACM Web Conference 2022, pp 2719--2728

    Fabbri F, Wang Y, Bonchi F, et al (2022) Rewiring what-to-watch-next recommendations to reduce radicalization pathways. In: Proceedings of the ACM Web Conference 2022, pp 2719--2728

    2022

  7. [15]

    In: Pacific Rim Knowledge Acquisition Workshop, Springer, pp 134--143

    Febrinanto FG, Liu M, Xia F (2023 a ) Balanced graph structure information for brain disease detection. In: Pacific Rim Knowledge Acquisition Workshop, Springer, pp 134--143

    2023

  8. [16]

    arXiv preprint arXiv:231209478

    Febrinanto FG, Moore K, Thapa C, et al (2023 b ) Entropy causal graphs for multivariate time series anomaly detection. arXiv preprint arXiv:231209478

    2023

  9. [17]

    IEEE Computational Intelligence Magazine 18(1):32--51

    Febrinanto FG, Xia F, Moore K, et al (2023 c ) Graph lifelong learning: A survey. IEEE Computational Intelligence Magazine 18(1):32--51

    2023

  10. [18]

    Advances in neural information processing systems 32

    Gasteiger J, Wei enberger S, G \"u nnemann S (2019) Diffusion improves graph learning. Advances in neural information processing systems 32

    2019

  11. [19]

    Econometrica: journal of the Econometric Society pp 424--438

    Granger CW (1969) Investigating causal relations by econometric models and cross-spectral methods. Econometrica: journal of the Econometric Society pp 424--438

    1969

  12. [20]

    Advances in neural information processing systems 35:8291--8303

    Han K, Wang Y, Guo J, et al (2022) Vision gnn: An image is worth graph of nodes. Advances in neural information processing systems 35:8291--8303

    2022

  13. [21]

    In: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE, pp 2809--2816

    Hu J, Huang Y, Dong S (2023) Transformer and snowball graph convolution learning for brain functional network analysis. In: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE, pp 2809--2816

    2023

  14. [22]

    Neuroimage 80:360--378

    Hutchison RM, Womelsdorf T, Allen EA, et al (2013) Dynamic functional connectivity: promise, issues, and interpretations. Neuroimage 80:360--378

    2013

  15. [23]

    IEEE Transactions on Industrial Informatics 17(9):6347--6356

    Jiao J, Zhen W, Zhu W, et al (2020) Quality-related root cause diagnosis based on orthogonal kernel principal component regression and transfer entropy. IEEE Transactions on Industrial Informatics 17(9):6347--6356

    2020

  16. [24]

    In: International Conference on Medical Imaging with Deep Learning, PMLR, pp 618--637

    Kan X, Cui H, Lukemire J, et al (2022 a ) Fbnetgen: Task-aware gnn-based fmri analysis via functional brain network generation. In: International Conference on Medical Imaging with Deep Learning, PMLR, pp 618--637

    2022

  17. [25]

    Advances in Neural Information Processing Systems 35:25586--25599

    Kan X, Dai W, Cui H, et al (2022 b ) Brain network transformer. Advances in Neural Information Processing Systems 35:25586--25599

    2022

  18. [26]

    NeuroImage 146:1038--1049

    Kawahara J, Brown CJ, Miller SP, et al (2017) Brainnetcnn: Convolutional neural networks for brain networks; towards predicting neurodevelopment. NeuroImage 146:1038--1049

    2017

  19. [27]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2):1606--1617

    Kazi A, Cosmo L, Ahmadi SA, et al (2022) Differentiable graph module (dgm) for graph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2):1606--1617

    2022

  20. [28]

    Human Brain Mapping 44(15):5167--5179

    Khalilullah KI, Agcaoglu O, Sui J, et al (2023) Multimodal fusion of multiple rest fmri networks and mri gray matter via parallel multilink joint ica reveals highly significant function/structure coupling in alzheimer's disease. Human Brain Mapping 44(15):5167--5179

    2023

  21. [29]

    European Radiology 31(7):5251--5252

    Kocak B (2021) Artificial intelligence to predict task activation from resting state fmri. European Radiology 31(7):5251--5252

    2021

  22. [30]

    arXiv preprint arXiv:240700984

    Li C, Yu S, Cui Y (2024 a ) Individual brain parcellation: Review of methods, validations and applications. arXiv preprint arXiv:240700984

    2024

  23. [31]

    In: 2024 International Joint Conference on Neural Networks (IJCNN), IEEE, pp 1--8

    Li H, Li C, Zhang J, et al (2024 b ) Addressing over-squashing in gnns with graph rewiring and ordered neurons. In: 2024 International Joint Conference on Neural Networks (IJCNN), IEEE, pp 1--8

    2024

  24. [32]

    Medical Image Analysis 74:102233

    Li X, Zhou Y, Dvornek N, et al (2021) Braingnn: Interpretable brain graph neural network for fmri analysis. Medical Image Analysis 74:102233

    2021

  25. [33]

    In: Australasian Joint Conference on Artificial Intelligence, Springer, pp 376--389

    Liu M, Dong Q, Wang C, et al (2024) Motif-induced subgraph generative learning for explainable neurological disorder detection. In: Australasian Joint Conference on Artificial Intelligence, Springer, pp 376--389

    2024

  26. [34]

    Schizophrenia research 83(2-3):155--171

    Makris N, Goldstein JM, Kennedy D, et al (2006) Decreased volume of left and total anterior insular lobule in schizophrenia. Schizophrenia research 83(2-3):155--171

    2006

  27. [35]

    Annual Review of Neuroscience 44:315--334

    Philiastides MG, Tu T, Sajda P (2021) Inferring macroscale brain dynamics via fusion of simultaneous eeg-fmri. Annual Review of Neuroscience 44:315--334

    2021

  28. [36]

    arXiv preprint arXiv:240814811

    Ren J, Xia F (2024) Brain-inspired artificial intelligence: A comprehensive review. arXiv preprint arXiv:240814811

    2024

  29. [37]

    Physical review letters 85(2):461

    Schreiber T (2000) Measuring information transfer. Physical review letters 85(2):461

    2000

  30. [38]

    Advances in Neural Information Processing Systems 37:94568--94606

    Tieu K, Fu D, Zhu Y, et al (2025) Temporal graph neural tangent kernel with graphon-guaranteed. Advances in Neural Information Processing Systems 37:94568--94606

    2025

  31. [39]

    In: International Conference on Learning Representations

    Topping J, Di Giovanni F, Chamberlain BP, et al (2022) Understanding over-squashing and bottlenecks on graphs via curvature. In: International Conference on Learning Representations

    2022

  32. [40]

    Neurobiology of Disease 181:106095

    Vasilkovska T, Adhikari MH, Van Audekerke J, et al (2023) Resting-state fmri reveals longitudinal alterations in brain network connectivity in the zq175dn mouse model of huntington's disease. Neurobiology of Disease 181:106095

    2023

  33. [41]

    In: ICLR

    Welling M, Kipf TN (2017) Semi-supervised classification with graph convolutional networks. In: ICLR

    2017

  34. [42]

    IEEE Transactions on Artificial Intelligence 2(2):109--127

    Xia F, Sun K, Yu S, et al (2021) Graph learning: A survey. IEEE Transactions on Artificial Intelligence 2(2):109--127

    2021

  35. [43]

    ACM Transactions on Intelligent Systems and Technology (TIST) 13(1):1--54

    Zhou Y, Zheng H, Huang X, et al (2022) Graph neural networks: Taxonomy, advances, and trends. ACM Transactions on Intelligent Systems and Technology (TIST) 13(1):1--54

    2022

  36. [44]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished institution journal key keywords month note number organization pages publisher school series title type url volume year archivePrefix primaryClass adsurl adsnote version label extra.labe...

  37. [45]

    write newline

    " write newline "" before.all 'output.state := FUNCTION add.period duplicate empty 'skip "." * add.blank if FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap dupl...

Pith tools

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