Pith. sign in

REVIEW 3 major objections 6 minor 25 references

An SDE-driven hypergraph network recovers continuous brain trajectories from irregular fMRI visits and improves Alzheimer's progression prediction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 21:32 UTC pith:3CPM3XHI

load-bearing objection Solid compositional methods paper for irregular longitudinal fMRI: real gains on OASIS-3/ADNI, but attribution of those gains to the joint SDE+hypergraph design is only partly isolated. the 3 major comments →

arxiv 2603.20452 v2 pith:3CPM3XHI submitted 2026-03-20 cs.LG

SDE-Driven Spatio-Temporal Hypergraph Neural Networks for Irregular Longitudinal fMRI Connectome Modeling in Alzheimer's Disease

Ruiying Chen , Yutong Wang , Houliang Zhou , Wei Liang , Yong Chen , Lifang He This is my paper
classification cs.LG
keywords Alzheimer's diseaselongitudinal fMRIirregular samplingstochastic differential equationshypergraph neural networksfunctional connectomespatio-temporal modelinginterpretability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Clinical brain scans for Alzheimer's disease arrive at uneven intervals, with missing visits and noisy region-level signals, so standard graph models that assume fixed time steps and only pairwise connections struggle to track progression. This paper argues that you can recover continuous latent trajectories of each brain region with a neural stochastic differential equation, then build dynamic hypergraphs that link multiple regions at once, and finally let the hypergraph convolution weights themselves evolve by an SDE conditioned on the real time gaps between visits. A sparsity regularizer further highlights which regions and multi-region connections matter most. On two public cohorts the resulting model beats existing graph and hypergraph baselines for predicting who will progress and for diagnostic classification, and the learned importance maps track known early and later stages of Alzheimer's network disruption.

Core claim

SDE-HGNN is presented as the first framework that jointly models continuous-time longitudinal dynamics and higher-order multi-region interactions for irregular fMRI connectomes; by reconstructing ROI signals as SDE trajectories, constructing KNN hypergraphs, evolving convolution parameters with inter-visit intervals, and sparsifying for interpretability, it yields consistent gains over graph and hypergraph baselines on OASIS-3 progression prediction and ADNI diagnosis.

What carries the argument

SDE-HGNN: neural SDEs first reconstruct continuous ROI trajectories, dynamic KNN hypergraphs capture higher-order connectivity, and SDE-controlled recurrent updates evolve hypergraph convolution weights according to real inter-visit time gaps, with a sparsity module selecting salient regions and hyperedges.

Load-bearing premise

The claim rests on the premise that KNN hyperedges from reconstructed ROI features plus SDE evolution of weights by inter-visit time gaps faithfully capture true higher-order functional interactions and subject-specific disease tempo, rather than gains coming mainly from extra capacity or smoothing.

What would settle it

If matched-capacity ablations that keep the SDE reconstruction and sparsity but replace continuous SDE weight evolution with discrete RNN or fixed-grid updates, or replace hyperedges with ordinary pairwise graphs, erase the reported AUC advantage on OASIS-3 as the number of irregular visits grows, the joint continuous higher-order design is not doing the claimed work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Irregular longitudinal fMRI can be treated as continuous stochastic trajectories rather than forced onto fixed visit grids.
  • Hypergraph convolution parameters can adapt to each subject's actual inter-scan intervals for disease-stage-aware connectivity modeling.
  • Sparsity over hypernodes and hyperedges can surface time-evolving ROI and multi-region biomarkers that track known Alzheimer's progression patterns.
  • The same pipeline can be applied to both multi-visit progression prediction and cross-sectional diagnostic tasks with shared continuous higher-order machinery.

Where Pith is reading between the lines

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

  • The same SDE-plus-hypergraph pattern could transfer to other irregular longitudinal biomarkers (PET, EEG, wearable streams) where visits are sparse and multi-region coordination matters.
  • If the continuous weight evolution is the true driver, clinics could schedule uneven follow-ups without sacrificing model accuracy, reducing burden on patients who miss visits.
  • Learned hyperedge importance that already separates DAN–DMN early and later prefrontal/temporo-parietal shifts suggests a natural route to subject-level progression staging without extra labels.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes SDE-HGNN for irregular longitudinal rs-fMRI connectome modeling in Alzheimer’s disease. It reconstructs ROI BOLD series with a neural SDE encoder–decoder (Eq. 1), builds per-visit KNN hypergraphs from Euclidean distances of reconstructed features (Eq. 2, K=10), evolves hypergraph convolution weights via an SDE conditioned on inter-visit Δt with GRU refinement (Eqs. 5–7), and applies sparsity-based node/hyperedge importance (Eqs. 8–11) with CE + MI + ℓ1 + entropy losses (Eq. 12). On OASIS-3 progressive-vs-stable prediction under 1–6 visits and three ADNI binary diagnostic tasks, the method reports higher AUC than cross-sectional graph/hypergraph and longitudinal baselines (including SDEGCN), with ablations of temporal module and sparsity (Table 3) and ROI/hyperedge visualizations aligned with known AD patterns.

Significance. If the gains are truly driven by joint continuous-time and higher-order modeling under irregular sampling, the work is a useful contribution at the intersection of neural SDEs, hypergraph learning, and longitudinal neuroimaging. Strengths include subject-level five-fold CV, evaluation across visit counts on OASIS-3, three ADNI tasks, ablations of SDE vs RNN/ODE and sparsity, hyperparameter sensitivity, and publicly linked code. The clinical interpretability analysis (parahippocampal early salience; DAN–DMN involvement) is a plus if the attribution of performance is tightened. The main open issue is causal isolation of the claimed design, not the empirical setup itself.

major comments (3)
  1. [Tables 1–3; Methodology (Eqs. 1–7)] Central claim (abstract/contributions; Tables 1–2) attributes AUC gains to jointly modeling continuous-time dynamics and higher-order interactions. Table 3 only swaps the temporal evolution (RNN/ODE vs SDE) or removes sparsity, while keeping SDE reconstruction and KNN hypergraphs fixed. There is no control that feeds the same reconstructed trajectories into a pairwise GNN (e.g., SDEGCN-style) or that runs the full stack on non-reconstructed irregular series. Without those isolations, gains—especially growth with visit count vs SDEGCN—could come from reconstruction smoothing, dual-SDE+GRU capacity, or sparsity rather than the claimed joint design. A load-bearing revision is to add at least one such control and report AUC under the same splits.
  2. [Hypergraph Construction; Eq. (2)] Hypergraph construction (Eq. 2) defines hyperedges via KNN on Euclidean distances of reconstructed ROI features, not on multi-region co-activation or correlation structure. The manuscript does not show that these hyperedges capture true higher-order functional interactions beyond denser pairwise neighborhoods. A short validation (e.g., comparison to correlation- or mutual-information-based hyperedges, or sensitivity of AUC to K and distance metric) is needed to support the higher-order interpretation of the performance lift.
  3. [Sparse Interpretability; Figs. 2–3; Table 3] Sparsity is presented both as an interpretability tool and as improving prediction (Table 3: full vs w/o sparsity). The clinical claims (Figs. 2–3: parahippocampal early salience; DAN–DMN connections) rest on learned importance probabilities and post-hoc t-tests, but there is no quantitative check that selected ROIs/hyperedges are stable across folds or predictive when used alone. Without stability or held-out biomarker validation, the biomarker narrative remains qualitative and should be tempered or strengthened with fold-wise consistency metrics.
minor comments (6)
  1. [Abstract vs Methodology] Notation inconsistency: abstract says “inter-scan intervals” while the body uses inter-visit Δtk; unify terminology.
  2. [Eq. (6)] Eq. (6) adds two GRU terms (current features with evolved weights and with W0); a one-sentence justification for the residual baseline would help.
  3. [Table 1] Table 1: for one time point, SDE-HGNN sensitivity is low (0.23) despite high specificity; discuss class imbalance or threshold choice.
  4. [Title block / affiliations] Author name formatting (“Y ong Chen”, “New Y ork”) and occasional spacing artifacts should be cleaned.
  5. [SDE-based reconstruction; Loss Function] Clarify whether SDE reconstruction is trained jointly end-to-end with the classifier or as a separate stage, and whether reconstruction loss is included in Eq. (12).
  6. [Hyperparameter Sensitivity Analysis] Fig. 4 reports inverted-U sensitivity for λ1–λ3; state the fixed values of other hyperparameters and the exact fold protocol for that sweep.

Circularity Check

0 steps flagged

No significant circularity: empirical architecture paper whose claims rest on held-out CV metrics against external baselines, not on definitional or self-justifying reductions.

full rationale

SDE-HGNN is a proposed neural architecture (SDE reconstruction of ROI series, KNN hypergraphs from reconstructed features, SDE-evolved hypergraph weights conditioned on Δt, plus sparsity/MI regularizers) evaluated by five-fold subject-level CV on public OASIS-3 and ADNI cohorts. Reported AUC/accuracy gains (Tables 1–2) and ablation drops (Table 3) are ordinary empirical outcomes of training the composite loss (Eq. 12) and measuring held-out classification; they are not forced by construction from any fitted parameter, self-definition, or uniqueness theorem. The SDE reconstruction (Eq. 1) and hypergraph incidence (Eq. 2) are modeling choices whose fidelity is tested, not assumed tautological. Prior SDE-GNN work (ref. 8) is cited as a related baseline and inspiration for continuous-time dynamics, but the paper’s central novelty claim (joint continuous-time + higher-order modeling) and performance numbers do not reduce to that citation; the same holds for the GNNExplainer-inspired sparsity mechanism. No equation equates a “prediction” to an input by definition, no uniqueness result is imported to forbid alternatives, and no known empirical pattern is merely renamed. Attribution gaps in the ablations (reconstruction vs. hypergraph vs. capacity) are experimental-design issues, not circularity of the derivation chain. The paper is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on modeling choices and tuned knobs rather than on a parameter-free theory. Load-bearing free parameters include KNN size, loss weights, and architecture widths. Domain axioms include that Euclidean KNN hyperedges capture higher-order functional organization and that SDEs are appropriate for both intra-scan BOLD and inter-visit weight dynamics. Invented entities are architectural constructs (SDE-evolved hypergraph weights, the full SDE-HGNN stack), not new physical objects.

free parameters (5)
  • K (KNN hyperedge size)
    Fixed to 10 for hyperedge construction; directly defines higher-order neighborhoods and thus the spatial inductive bias.
  • λ1, λ2, λ3 loss weights
    Set to 2, 0.1, 0.1 after sensitivity search; control MI, L1 sparsity, and entropy terms that shape both accuracy and interpretability.
  • SDE drift/diffusion network parameters and latent dimension
    Learned neural SDE encoder–decoder parameters determine reconstructed ROI trajectories that all downstream connectomes depend on.
  • Hypergraph convolution hidden dim dl, GRU/SDE weight-state size, learning rate
    Selected on validation folds; capacity and optimization choices affect reported margins over baselines.
  • q distance scaling in incidence weights
    Scaling parameter in Eq. (2) modulates soft hyperedge membership; not derived from first principles.
axioms (5)
  • domain assumption Higher-order multi-region interactions relevant to AD are adequately represented by KNN hyperedges on Euclidean ROI feature distances.
    Hypergraph Construction section replaces correlation graphs with distance-based KNN hyperedges without independent validation that these hyperedges match true multi-region functional assemblies.
  • domain assumption Intra-scan BOLD irregularity and inter-visit disease progression are well modeled by neural SDEs (drift + Brownian diffusion).
    Eqs. (1) and (5) assume continuous stochastic dynamics are the right generative form; ablation vs ODE/RNN supports utility but not uniqueness or biological correctness.
  • domain assumption Subject-level stratified 5-fold CV on OASIS-3 progressive vs stable and ADNI binary labels is a valid estimate of clinical generalization.
    Experiments section; no external multi-site locked test set is reported.
  • standard math Standard neural SDE / hypergraph convolution / GRU / GNNExplainer-style math can be composed as written.
    Uses established SDE, incidence-matrix normalization, and GRU update forms from cited literature.
  • ad hoc to paper Sparsity probabilities on nodes/hyperedges both improve prediction and yield clinically meaningful biomarkers.
    Sparse Interpretability and loss terms Ls/Le/Lmi are design choices; ablation shows some gain, but clinical meaning is interpretive.
invented entities (2)
  • SDE-HGNN architecture no independent evidence
    purpose: Joint continuous-time reconstruction, dynamic hypergraph connectomes, and SDE-evolved hypergraph convolution for irregular longitudinal fMRI.
    The stacked system is the paper’s main proposed object; evidence is internal empirical comparison, not an external physical discovery.
  • SDE-controlled hypergraph convolution weight process W_tk no independent evidence
    purpose: Let filter parameters evolve with inter-visit intervals then be GRU-refined for disease-stage-adaptive connectivity modeling.
    Eqs. (5)–(7) introduce this continuous weight state; no independent measurement of such weights outside the model.

pith-pipeline@v1.1.0-grok45 · 16178 in / 3761 out tokens · 44626 ms · 2026-07-13T21:32:39.325036+00:00 · methodology

0 comments
read the original abstract

Longitudinal neuroimaging is essential for modeling disease progression in Alzheimer's disease (AD), yet irregular sampling and missing visits pose substantial challenges for learning reliable temporal representations. To address this challenge, we propose SDE-HGNN, a stochastic differential equation (SDE)-driven spatio-temporal hypergraph neural network for irregular longitudinal fMRI connectome modeling. The framework first employs an SDE-based reconstruction module to recover continuous latent trajectories from irregular observations. Based on these reconstructed representations, dynamic hypergraphs are constructed to capture higher-order interactions among brain regions over time. To further model temporal evolution, hypergraph convolution parameters evolve through SDE-controlled recurrent dynamics conditioned on inter-visit intervals, enabling disease-stage-adaptive connectivity modeling. We also incorporate a sparsity-based importance learning mechanism to identify salient brain regions and discriminative connectivity patterns. Extensive experiments on the OASIS-3 and ADNI cohorts demonstrate consistent improvements over state-of-the-art graph and hypergraph baselines in AD progression prediction. The source code is available at https://anonymous.4open.science/r/SDE-HGNN-017F.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references

  1. [1]

    Hypothetical model of dynamic biomarkers of the Alzheimer’s pathological cascade

    Jack CR, Knopman DS, Jagust WJ, Shaw LM, Aisen PS, Weiner MW, et al. Hypothetical model of dynamic biomarkers of the Alzheimer’s pathological cascade. The Lancet Neurology. 2010;9(1):119-28

    2010

  2. [2]

    What we can do and what we cannot do with fMRI

    Logothetis NK. What we can do and what we cannot do with fMRI. Nature. 2008;453(7197):869-78

    2008

  3. [3]

    Overview of fMRI analysis

    Smith S. Overview of fMRI analysis. The British journal of radiology. 2004;77:S167-75

    2004

  4. [4]

    Graph neural networks in network neuroscience

    Bessadok A, Mahjoub MA, Rekik I. Graph neural networks in network neuroscience. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2022;45(5):5833-48

    2022

  5. [5]

    fMRI-based brain disease diagnosis: A graph network approach

    Tong W, Li YX, Zhao XY , Chen QQ, Gao YB, Li P, et al. fMRI-based brain disease diagnosis: A graph network approach. IEEE Transactions on Medical Robotics and Bionics. 2023;5(2):312-22

    2023

  6. [6]

    Beyond the snapshot: Brain tokenized graph transformer for longitudinal brain functional connectome embedding

    Dong Z, Wu Y , Xiao Y , Chong JSX, Jin Y , Zhou JH. Beyond the snapshot: Brain tokenized graph transformer for longitudinal brain functional connectome embedding. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer; 2023. p. 348-57

    2023

  7. [7]

    Brainode: Dynamic brain signal analysis via graph- aided neural ordinary differential equations

    Han K, Yang Y , Huang Z, Kan X, Guo Y , Yang Y , et al. Brainode: Dynamic brain signal analysis via graph- aided neural ordinary differential equations. In: IEEE EMBS International Conference on Biomedical and Health Informatics. IEEE; 2024. p. 1-8

    2024

  8. [8]

    Uncovering alzheimer’s disease progression via SDE- based spatio-temporal graph deep learning on longitudinal brain networks

    Zhou H, Zhou R, Liu Y , Zhao K, Shen L, Chen BY , et al. Uncovering alzheimer’s disease progression via SDE- based spatio-temporal graph deep learning on longitudinal brain networks. arXiv preprint arXiv:250921735. 2025

    2025

  9. [9]

    Modelling functional brain aging in Alzheimer’s disease using neural ODEs on rsfMRI graphs

    Mahmud AI, Suleiman AB, Yahaya U. Modelling functional brain aging in Alzheimer’s disease using neural ODEs on rsfMRI graphs. International Journal of Research Publication and Reviews. 2025;6(7):2482-8

    2025

  10. [10]

    Dynamic weighted hypergraph convolutional network for brain functional connectome analysis

    Wang J, Li H, Qu G, Cecil KM, Dillman JR, Parikh NA, et al. Dynamic weighted hypergraph convolutional network for brain functional connectome analysis. Medical image analysis. 2023;87:102828

    2023

  11. [11]

    Hypergale: Asd classification via hypergraph gated attention with learnable hyperedges

    Arora M, Jain CS, Baru LB, Dadi K, et al. Hypergale: Asd classification via hypergraph gated attention with learnable hyperedges. In: 2024 International Joint Conference on Neural Networks (IJCNN). IEEE; 2024. p. 1-8

    2024

  12. [12]

    Hypergraph-based semantic and topological self-supervised learning for brain disease diag- nosis

    Han X, Lei M, Li J. Hypergraph-based semantic and topological self-supervised learning for brain disease diag- nosis. Pattern Recognition. 2026;169:111921

    2026

  13. [13]

    Hypergraph convolutional network for longitudinal data analysis in Alzheimer’s disease

    Hao X, Li J, Ma M, Qin J, Zhang D, Liu F, et al. Hypergraph convolutional network for longitudinal data analysis in Alzheimer’s disease. Computers in Biology and Medicine. 2024;168:107765

    2024

  14. [14]

    GRU-ODE-Bayes: Continuous modeling of sporadically-observed time series

    De Brouwer E, Simm J, Arany A, Moreau Y . GRU-ODE-Bayes: Continuous modeling of sporadically-observed time series. Advances in neural information processing systems. 2019;32

    2019

  15. [15]

    Gnnexplainer: Generating explanations for graph neural networks

    Ying Z, Bourgeois D, You J, Zitnik M, Leskovec J. Gnnexplainer: Generating explanations for graph neural networks. Advances in neural information processing systems. 2019;32

    2019

  16. [16]

    Open access series of imaging studies: longitu- dinal MRI data in nondemented and demented older adults

    Marcus DS, Fotenos AF, Csernansky JG, Morris JC, Buckner RL. Open access series of imaging studies: longitu- dinal MRI data in nondemented and demented older adults. Journal of cognitive neuroscience. 2010;22(12):2677- 84

    2010

  17. [17]

    The Alzheimer’s disease neuroimaging initiative

    Mueller SG, Weiner MW, Thal LJ, Petersen RC, Jack C, Jagust W, et al. The Alzheimer’s disease neuroimaging initiative. Neuroimaging Clinics. 2005;15(4):869-77

    2005

  18. [18]

    fMRIPrep: a robust preprocessing pipeline for functional MRI

    Esteban O, Markiewicz CJ, Blair RW, et al. fMRIPrep: a robust preprocessing pipeline for functional MRI. Nature Methods. 2019;16(1):111-6

    2019

  19. [19]

    ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data

    Pruim RH, Mennes M, van Rooij D, et al. ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data. NeuroImage. 2015;112:267-77

    2015

  20. [20]

    Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity

    Schaefer A, Kong R, Gordon EM, Laumann TO, Zuo XN, Holmes AJ, et al. Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity. Cerebral cortex. 2018;28(9):3095-114

    2018

  21. [21]

    A dynamic graph convolutional neural network frame- work reveals new insights into connectome dysfunctions in ADHD

    Zhao K, Duka B, Xie H, Oathes DJ, Calhoun V , Zhang Y . A dynamic graph convolutional neural network frame- work reveals new insights into connectome dysfunctions in ADHD. NeuroImage. 2022;246:118774

    2022

  22. [22]

    Atrophy in the parahippocampal gyrus as an early biomarker of Alzheimer’s disease

    Echávarri C, Aalten P, Uylings HB, Jacobs H, Visser PJ, Gronenschild E, et al. Atrophy in the parahippocampal gyrus as an early biomarker of Alzheimer’s disease. Brain Structure and Function. 2011;215(3):265-71

    2011

  23. [23]

    The parahippocampal gyrus in Alzheimer’s disease: clinical and preclinical neuroanatomical correlates

    Van Hoesen GW, Augustinack JC, Dierking J, Redman SJ, Thangavel R. The parahippocampal gyrus in Alzheimer’s disease: clinical and preclinical neuroanatomical correlates. Annals of the new York Academy of Sciences. 2000;911(1):254-74

    2000

  24. [24]

    Staging of Alzheimer disease- associated neurofibrillary pathology using paraffin sections and immunocytochemistry

    Braak H, Alafuzoff I, Arzberger T, Kretzschmar H, Del Tredici K. Staging of Alzheimer disease- associated neurofibrillary pathology using paraffin sections and immunocytochemistry. Acta neuropathologica. 2006;112(4):389-404

    2006

  25. [25]

    Four distinct trajectories of tau deposition identified in Alzheimer’s disease

    V ogel JW, Young AL, Oxtoby NP, Smith R, Ossenkoppele R, Strandberg OT, et al. Four distinct trajectories of tau deposition identified in Alzheimer’s disease. Nature medicine. 2021;27(5):871-81

    2021