Pith. sign in

REVIEW 4 major objections 5 minor 41 references

A Stage-Aware Mixture of Experts Framework for Neurodegenerative Disease Progression Modelling

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A mixture-of-experts model that lets mechanism weights shift over disease stage predicts tau accumulation better and reveals graph spread matters early, local processes later.

desk verdict A useful hybrid MoE for tau progression with a defensible predictive core, but the stage-weight mechanistic claims outrun the evidence. read the letter →

arxiv 2508.07032 v1 pith:DIASVDEO submitted 2025-08-09 cs.LG q-bio.QM

Tiantian He , Keyue Jiang , An Zhao , Anna Schroder , Elinor Thompson , Sonja Soskic , Frederik Barkhof , Daniel C. Alexander This is my paper · ORCID
classification cs.LGq-bio.QM
keywords taupathologydiseaseprogressionmodellingmixtureofexpertsgraphneuraldiffusionODEtemporalalignmentAlzheimer'slongitudinalsparsedata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the long-term spread of tau protein in Alzheimer's disease is better described by a stage-dependent mixture of mechanisms than by any single fixed mechanism. The proposed model writes the rate of change of pathology as a sum of three experts—a classical pathophysiological diffusion-reaction term, an inhomogeneous graph neural diffusion term, and a localized neural reaction term—with weights $\beta_1(t), \beta_2(t), \beta_3(t)$ that evolve over the disease course. A dual optimization alternately places each subject's sparse, irregular scans on a common cohort timeline and refines that timeline with the mixture. If the claim is right, small longitudinal datasets can be used to reconstruct long-term trajectories, and the learned weights become quantitative hypotheses about which biological process dominates at each disease stage. The paper specifically claims this beats both purely physical and purely neural baselines on tau-PET prediction, with graph-driven spread strongest early and unknown local processes dominant later.

What carries the argument

The load-bearing object is the stage-aware mixture-of-experts ODE $$\frac{dc(t)}{dt} = \beta_1(t) f_M + \beta_2(t) f_S + \beta_3(t) f_L, \quad \sum_j \beta_j(t) = 1,$$ where $f_M$ is the pathophysiological network-diffusion plus logistic-reaction expert, $f_S$ is an inhomogeneous graph diffusion implemented by a graph auto-encoder that learns a refined adjacency from node states over time, and $f_L$ is an MLP capturing localized dynamics beyond reaction terms. The $\beta_j(t)$ are temporal attention weights. Two regularizers—a norm term that keeps the neural corrections subordinate to the physical expert and an orthogonality term that decorrelates the experts' outputs—are what allow the weig

What would settle it

Train the same mixture on synthetic trajectories generated with known, time-varying mixture weights $\beta_j(t)$ and check whether the estimated weights recover the planted curves; if the late-stage neural weight rises even when the true late mechanism is graph diffusion, the stage-wise mechanism interpretation is an artifact of residual-fitting.

Watch

Extended reading notes

Core claim

The central claim is that tau propagation through the brain is not governed by one fixed mechanism over the whole disease course. On the cohort level it is written as a weighted sum of three experts, each proposing a rate of tau change: a classical graph-diffusion plus local-reaction model, an inhomogeneous graph neural diffusion whose diffusivity depends on current pathology and time, and a localized neural reaction network. The weights $\beta_1(t), \beta_2(t), \beta_3(t)$ are learned and vary with stage, with the pathophysiological expert preferred when it fits and the neural experts used to capture the rest. The paper reports that this mixture beats the physical model and all compared neu

Load-bearing premise

The clinical interpretation rests on the assumption that, after the two regularizers, the three experts' outputs are identifiable and biologically separable, so the later-stage dominance of the flexible neural term is a discovered unknown process rather than the network absorbing everything the physical terms cannot fit.

Editorial extensions

If this is right

  • Long-term disease trajectories can be reconstructed from short, irregularly spaced individual scans as long as enough subjects cover different stages, easing a core data bottleneck in neurodegeneration research.
  • A mechanism's contribution to tau progression should not be treated as constant; models that fix the mixing weights lose prediction accuracy and misstate which process matters when.
  • The learned mixture weights yield a stage-specific biological hypothesis: connectivity-driven spread dominates early tau accumulation, while non-graph local processes dominate later, consistent with prior work on local seed accumulation.
  • The inhomogeneous graph diffusion component indicates that a fixed structural connectome understates network dynamics; refining the propagation graph from node states improves long-term forecasts over homogeneous diffusion models.
  • Combining mechanistic and neural components under a constrained mixture is more robust for small datasets than using either purely mechanistic or purely data-driven models.

Reading between the lines

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

  • Editorial inference: the 'graph early, unknown later' story is only as trustworthy as the identifiability of the three experts; a natural next check is a synthetic recovery experiment with planted time-varying weights.
  • Editorial inference: because the pipeline itself does not use disease-specific parameters beyond the physical expert, the same dual-alignment plus stage-weighted mixture could be applied to other proteinopathies or biomarkers.
  • Editorial inference: replacing the late-stage MLP with named biological mechanisms, such as clearance, inflammation, or glial response, would turn the residual 'unknown local process' into a testable set of hypotheses.
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 / 5 minor

Summary. The manuscript proposes IGND-MoE, a stage-aware mixture-of-experts framework for modelling long-term neurodegenerative disease progression from sparse, irregular longitudinal observations. The trajectory derivative is a time-weighted sum of three experts: the classical pathophysiological Fisher--Kolmogorov model f_M, an inhomogeneous graph neural diffusion model f_S implemented with a graph auto-encoder (GAE), and an unconstrained MLP-based localized reaction model f_L. Temporal weights beta_j(t) (summing to one) are learned, and the cohort trajectory and individual subject time offsets are estimated by iterative dual optimization. On tau-PET data from ADNI (216 subjects, 378 observations), the model reports improved held-out prediction (SSE 13.97 +/- 0.30, R 0.717 +/- 0.014) over several physical, discrete, and neural-ODE baselines. The authors further interpret the learned stage weights as clinical evidence that graph-related processes dominate early and 'other unknown physical processes' dominate later, a claim stated in the abstract and Section 4.3.

Significance. If the predictive and mechanistic claims are both established, the framework would be a useful hybrid model for disease progression modelling: it combines an interpretable physical prior with flexible neural components and yields testable, stage-dependent mechanism weights. Strengths of the paper include a clear problem formulation, inclusion of several baseline families, held-out evaluation, and explicit regularization intended to encourage expert complementarity. The predictive improvement in Table 1 is plausible and seems worth building on. However, the central mechanistic interpretation is not currently supported: the beta_j(t) weights are not shown to be identifiable as biologically separate mechanisms, and the trajectory is initialized from the same physical model that is one of the experts, creating a possible circularity. The external validation claims are also deferred to omitted appendices. The mechanistic half of the central claim therefore requires substantial additional support.

major comments (4)
  1. [Section 3.4.2, Eq. (3) and Eq. (10)] The stage weights beta_j(t) are not identifiable as contributions of distinct biological mechanisms. For any nonzero scalar s, replacing f_S by s*f_S and beta_2 by beta_2/s leaves the product beta_2*f_S unchanged, and similarly for f_L. The norm loss Lnorm and orthogonality loss Lortho in Eq. (10) only bound magnitudes and decorrelate outputs; they do not fix this scaling degeneracy, and decorrelation does not imply biological separation. Since f_L is an unconstrained MLP, it can absorb residual misfit, and its late-stage dominance may simply reflect capacity allocation. The statement in Sec. 4.3 that 'other unknown physical processes become dominant later' is therefore a restatement of the fitted weights, not a validated finding. Please provide an identifiability analysis (e.g., simulation with known ground-truth beta_j(t)) or external validation of the weight curves before making this
  2. [Section 3.2, step 1, and Section 3.3, Eq. (2)] The temporal alignment step initializes and repeatedly refines subject locations using the pathophysiological trajectory f_M, and f_M is also one of the experts in the mixture. This creates a potential circularity: early-stage dominance of beta_1 may be induced by the initialization rather than discovered from data. Please report sensitivity analyses, e.g., initializing the trajectory from a data-driven expert or from a random trajectory, and show how much the inferred beta_j(t) curves change. This is needed to support the claim that the weights reveal stage-dependent mechanisms rather than reflect the initialization.
  3. [Discussion and Appendix A.2] The Discussion states that the findings are 'validated across multiple independent datasets', and Section 4.1 refers to external validation on HABS and A4 as well as OOD robustness experiments in Appendices A.1--A.3. These appendices are not included in the submitted manuscript. As a result, the external-validation and generalization claims cannot be assessed. Please include the appendices or remove/weaken the claims. This is load-bearing for the robustness half of the central claim.
  4. [Section 3.4.1, Eqs. (5)--(8)] The derivation replaces the inhomogeneous diffusion operator K G(c(t),t) K^T by a low-rank GAE reconstruction (M_Dec M_Enc^T) and then, in Eq. (8), predicts the derivative as Dec(h(t), A_hat) with A_hat = sigma(h(t)h(t)^T). No formal or empirical evidence is given that this low-rank factorization approximates the original graph diffusion operator, nor is the relationship between Eq. (5) and Eq. (8) made explicit. Please provide a validation of the approximation (e.g., how close is the learned M_Dec M_Enc^T to K G K^T on the connectome?) or clarify the sense in which the GAE implements inhomogeneous graph diffusion.
minor comments (5)
  1. [Section 4.2, Table 1] The table is referenced as 'Table??' in the text. Also, the caption and text mention 'three metrics', but the table shows only SSE and R, with an undefined 'Corr' column. Please reconcile and define all metrics.
  2. [Eq. (10)] The definition of Lortho appears to contain a typo: lambda appears inside the sum and is also multiplied by lambda_2 outside. This makes the role of lambda_2 unclear. Please rewrite the loss definition.
  3. [Section 3.4.1, Eq. (3)] The weights beta_j(t) are said to sum to one, but no non-negativity constraint is stated. If the weights can be negative or larger than one, their interpretation as 'contributions' is unclear. Please specify the admissible range of beta_j(t).
  4. [Section 3.3, Eq. (2) and Section 3.4.2, Eq. (10)] The notation for the trajectory loss is inconsistent between L_fM in Eq. (2) and L_traj in Eq. (10). Please unify the notation and make explicit which trajectory is used in the alignment step at each iteration.
  5. [Section 3.2] There is a typo: 'Temporal Initalization Step' should be 'Temporal Initialization Step'. More substantively, Step 1 says the trajectory is 'simulated by the pathophysiological model', but the details of how f_M is fit to data before the joint optimization are not given.

Circularity Check

1 steps flagged · score 6.0 of 10

The mechanistic conclusion about stage-dependent dominance is a restatement of the fitted mixture weights β_j(t), not an independent result; quantitative prediction remains externally evaluated.

  1. self definitional [Eq. (3) (Sec. 3.4.1) and Sec. 4.3 ('Qualitative Results Analysis')]
    "The temporal attention, which represents the model’s contribution versus disease stage, shows the physical model and GAE contributions decreasing with disease progression, while the MLP becomes dominant at later stages. This suggests graph-related processes are more influential early in our cohort, while other unknown physical processes dominate later stages."

    In Eq. (3) the 'graph-related processes' enter only through β2(t) times the term labelled fS:Graph Neural Diffusion, and the 'unknown physical processes' enter only through β3(t) times the MLP term fL:Localized Neural Reaction. Hence the Sec. 4.3 'suggestion' is exactly the fitted temporal weight vector β(t) read back in words: β2 large early and β3 large late is, by definition, 'graph-related processes early, unknown processes later.' No independent measurement, identifiability theorem, or ground-truth simulation establishes that the trained experts correspond to distinct biological mechanisms; the norm and orthogonality losses in Eq. (10) only bound and decorrelate outputs and do not fix the scaling degeneracy β_j f_j = (β_j/s)(s f_j). The mechanistic insight is therefore a renamed fitte

full rationale

The paper's quantitative claim is largely self-contained: the held-out SSE/R comparison in Table 1 is evaluated on a test split and does not reduce to the training fit, so the prediction-accuracy contribution is not circular. The circularity is concentrated in the second half of the abstract and in Sec. 4.3. There, the 'novel clinical insight' about mechanism dominance is a restatement of the fitted temporal-attention weights β_j(t) introduced in Eq. (3). Because fS is labelled by construction as graph neural diffusion and fL is an unconstrained MLP, saying that β3(t) dominates late is identical to saying the MLP contributes more late; renaming that residual as 'other unknown physical processes' does not add independent evidence. The regularization losses in Eq. (10) are asserted to 'disentangle' expert contributions, but no identifiability analysis or simulation with known ground-truth weights is provided in the main text. The external datasets (Appendix A.2) validate prediction accuracy, not the biological identity of the experts. The paper also cites its own prior work (He et al. 2023) only for a limitation, not as load-bearing support, so that does not raise the score. Overall, the quantitative prediction is externally evaluated, but the central mechanistic conclusion reduces by definition to the fitted weights, giving a partial circularity score of 6.

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

The central claims rest mainly on the inherited Fisher-Kolmogorov hypothesis, the one-dimensional cohort-trajectory assumption, the GAE low-rank approximation, and the use of regularizers as identifiability proxies. The physical model parameters and the temporal attention weights are fitted on the same training data, which is why the mechanistic findings carry circularity burden.

free parameters (5)
  • k, alpha, v (physical model constants in Eq. 1) = not reported in main text
    Graph diffusion rate, local production rate, and plateau value; these are presumably fitted to training tau SUVR data but the values and fitting procedure are not given.
  • lambda1, lambda2 in Eq. (10) = not reported
    Regularization weights balancing trajectory loss, norm loss, and orthogonality loss; they directly affect the inferred stage weights and are chosen without reported justification.
  • Temporal attention parameters defining beta_j(t) = learned
    These fitted weights are the basis of the clinical finding, so the 'prediction' of stage-dependent mechanisms is a description of fitted values.
  • GAE hidden dimensionality / rank e' and architecture hyperparameters = not reported
    Capacity of the learned diffusion operator; no sensitivity analysis or theory justifying the choice.
  • Tau-positivity cutoff (mean of negative Gaussian component plus one standard deviation) = computed per region from data
    Defines the inclusion criterion for the N=216 cohort; affects which subjects and observations enter the analysis.
assumptions (5)
  • domain assumption A single cohort-level disease trajectory exists and each subject's observations can be mapped onto it by a scalar time offset t0.
    Invoked in Sections 3.1 and 3.3 (Eq. 2). Individual differences beyond time shift are treated as noise; the paper itself hints at possible subgroups in Section 4.3.
  • domain assumption Tau spread follows a Fisher-Kolmogorov reaction-diffusion process with graph Laplacian diffusion and logistic local growth (Eq. 1).
    This is the backbone of fM, used both as prior trajectory initialization and as an expert. It is a physical hypothesis inherited from Raj et al. 2012a and Meisl et al. 2021a.
  • ad hoc to paper The inhomogeneous diffusion operator K G K^T can be replaced by a low-rank GAE reconstruction without loss of relevant dynamics (Eqs. 5-8).
    No formal equivalence is given between the low-rank factorization and the GAE decoder; this is an architectural choice that is not derived or validated.
  • ad hoc to paper The norm and orthogonality losses in Eq. (10) are sufficient to disentangle the three expert mechanisms.
    The penalties encourage diversity but do not guarantee semantic separation into graph, local, and physical processes; no identifiability analysis is provided.
  • domain assumption The group-averaged structural connectome from 50 external subjects is a stable substrate for cohort-level tau dynamics.
    Used to construct graph A in Eqs. (1) and (8); group averaging may remove individual connectivity differences relevant to progression.
invented entities (2)
  • Other unknown physical processes that dominate late disease stages
    purpose: Explain the residual dynamics captured by the MLP expert after graph diffusion is discounted
    No measurement or mechanism is specified; this is a label for unexplained residual, yet it is presented as a clinical finding in the Abstract and Section 5.
  • Refined graph A_hat produced by the GAE
    purpose: Data-driven replacement for the structural connectome in the spatial expert
    The paper treats it as implicitly refining the group connectome (Section 5) but provides no validation that it corresponds to biological connectivity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Stage-Aware Mixture of Experts Framework for Neurodegenerative Disease Progression Modelling." pith.science (2026). https://pith.science/paper/DIASVDEO

@misc{pith2026250807032,
  author       = {Pith},
  title        = {Pith review of: A Stage-Aware Mixture of Experts Framework for Neurodegenerative Disease Progression Modelling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIASVDEO}},
  note         = {Machine review of arXiv:2508.07032}
}
read the original abstract

The long-term progression of neurodegenerative diseases is commonly conceptualized as a spatiotemporal diffusion process that consists of a graph diffusion process across the structural brain connectome and a localized reaction process within brain regions. However, modeling this progression remains challenging due to 1) the scarcity of longitudinal data obtained through irregular and infrequent subject visits and 2) the complex interplay of pathological mechanisms across brain regions and disease stages, where traditional models assume fixed mechanisms throughout disease progression. To address these limitations, we propose a novel stage-aware Mixture of Experts (MoE) framework that explicitly models how different contributing mechanisms dominate at different disease stages through time-dependent expert weighting.Data-wise, we utilize an iterative dual optimization method to properly estimate the temporal position of individual observations, constructing a co hort-level progression trajectory from irregular snapshots. Model-wise, we enhance the spatial component with an inhomogeneous graph neural diffusion model (IGND) that allows diffusivity to vary based on node states and time, providing more flexible representations of brain networks. We also introduce a localized neural reaction module to capture complex dynamics beyond standard processes.The resulting IGND-MoE model dynamically integrates these components across temporal states, offering a principled way to understand how stage-specific pathological mechanisms contribute to progression. The stage-wise weights yield novel clinical insights that align with literature, suggesting that graph-related processes are more influential at early stages, while other unknown physical processes become dominant later on.

Figures

Figures reproduced from arXiv: 2508.07032 by the authors.

Figure 1
Figure 1. Stage-aware Neurodegenerative Disease Progression Modeling with IGND-MoE. This figure demonstrates the proposed IGND-MoE framework for constructing a full disease progression process from snapshots, by iteratively carrying out the temporal alignment step for mapping each subject to the proper location on the time axis and the trajectory construction step of shaping a better trajectory through the proposed temporal-a… view at source ↗
Figure 2
Figure 2. Temporal alignment for long-term progression. This figure visualizes how the proposed framework uses the snapshots of the individual cross-sectional data or the short longitudinal data to construct the full disease progression trajectory. Each colour represents one brain region. The dots represent real observations. The dots connected with dashed lines represent the longitudinal observations from the same subject, w… view at source ↗
Figure 3
Figure 3. Model predictions. Each curve depicts tau-accumulation trajectories for one of 68 cortical regions, with dots marking individual observations. SUVR signals across brain regions, which captures the spatial distribution similarity of pathology. These individual correlations are then averaged to obtain a group-level measure of the model’s ability to reproduce region-specific pathology patterns independent of absolute m… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Regional Error mapping on the brain. The plot displays the distribution of the error pattern during disease progression from the real observations using the physical model and the MoE-GAE model. The colour bar, shared by all brain plots, displays the extent of error le…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages

  1. [1]

    Simulad: a dynamical model for personalized simulation and disease staging in alzheimer’s disease

    Cl \'e ment Abi Nader, Federica Ribaldi, Giovanni B Frisoni, Valentina Garibotto, Philippe Robert, Nicholas Ayache, and Marco Lorenzi. Simulad: a dynamical model for personalized simulation and disease staging in alzheimer’s disease. Neurobiology of Aging, 113: 0 73--83, 2022

    work page 2022

  2. [2]

    Communication dynamics in complex brain networks

    Andrea Avena-Koenigsberger, Bratislav Misic, and Olaf Sporns. Communication dynamics in complex brain networks. Nature reviews neuroscience, 19 0 (1): 0 17--33, 2018

    work page 2018

  3. [3]

    Synergy between amyloid- and tau in Alzheimer’s disease

    Marc Aurel Busche and Bradley T Hyman. Synergy between amyloid- and tau in Alzheimer’s disease . Nature neuroscience, 23 0 (10): 0 1183--1193, 2020. ISSN 1097-6256

    work page 2020

  4. [4]

    Carslaw and J.C

    H.S. Carslaw and J.C. Jaeger. Conduction of Heat in Solids. Oxford science publications. Clarendon Press, 1959. ISBN 9780198533689. URL https://books.google.co.uk/books?id=y20sAAAAYAAJ

    work page 1959

  5. [5]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural Ordinary Differential Equations . 6 2018. URL http://arxiv.org/abs/1806.07366

    arXiv 2018

  6. [6]

    Empirical evaluation of gated recurrent neural networks on sequence modeling

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014

    arXiv 2014

  7. [7]

    Desikan et al

    Rahul S. Desikan et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. NeuroImage, 31 0 (3): 0 968--980, July 2006. ISSN 1053-8119. doi:10.1016/J.NEUROIMAGE.2006.01.021. Publisher: Academic Press

    doi:10.1016/j.neuroimage.2006.01.021 2006

  8. [8]

    R. A. FISHER. The wave of advance of advantageous genes. Annals of Eugenics, 7 0 (4): 0 355--369, 1937

    work page 1937

Show all 41 references
  1. [9]

    An event-based model for disease progression and its application in familial alzheimer's disease and huntington's disease

    Hubert M Fonteijn, Marc Modat, Matthew J Clarkson, Josephine Barnes, Manja Lehmann, Nicola Z Hobbs, Rachael I Scahill, Sarah J Tabrizi, Sebastien Ourselin, Nick C Fox, et al. An event-based model for disease progression and its application in familial alzheimer's disease and h...

    2012

  2. [10]

    Modeling and inference of spatio-temporal protein dynamics across brain networks

    Sara Garbarino, Marco Lorenzi, and Alzheimer’s Disease Neuroimaging Initiative. Modeling and inference of spatio-temporal protein dynamics across brain networks. In International Conference on Information Processing in Medical Imaging, pages 57--69. Springer, 2019 a

    2019

  3. [11]

    Differences in topological progression profile among neurodegenerative diseases from imaging data

    Sara Garbarino, Marco Lorenzi, Neil P Oxtoby, Elisabeth J Vinke, Razvan V Marinescu, Arman Eshaghi, M Arfan Ikram, Wiro J Niessen, Olga Ciccarelli, Frederik Barkhof, et al. Differences in topological progression profile among neurodegenerative diseases from imaging data. Elife...

    2019

  4. [12]

    Godsil and Gordon F

    Christopher D. Godsil and Gordon F. Royle. Algebraic Graph Theory. Graduate texts in mathematics. Springer, 2001

    2001

  5. [13]

    Long short-term memory

    Alex Graves and Alex Graves. Long short-term memory. Supervised sequence labelling with recurrent neural networks, pages 37--45, 2012

    2012

  6. [14]

    A coupled-mechanisms modelling framework for neurodegeneration

    Tiantian He, Elinor Thompson, Anna Schroder, Neil P Oxtoby, Ahmed Abdulaal, Frederik Barkhof, and Daniel C Alexander. A coupled-mechanisms modelling framework for neurodegeneration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, page...

    2023

  7. [15]

    Relational representation learning for dynamic (knowledge) graphs: A survey

    Seyed Mehran Kazemi, Rishab Goel, Kshitij Jain, Ivan Kobyzev, Akshay Sethi, Peter Forsyth, and Pascal Poupart. Relational representation learning for dynamic (knowledge) graphs: A survey. CoRR, abs/1905.11485, 2019

    1905 arXiv

  8. [16]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR (Poster) . OpenReview.net, 2017

    2017

  9. [17]

    Flortaucipir ( AV -1451) processing methods

    Susan Landau, Tyler J Ward, Alice Murphy, and William Jagust. Flortaucipir ( AV -1451) processing methods. page 8, 2021

    2021

  10. [18]

    Comparison of group-level and individualized brain regions for measuring change in longitudinal tau positron emission tomography in alzheimer disease

    Antoine Leuzy, Alexa Pichet Binette, Jacob W Vogel, Gregory Klein, Edilio Borroni, Matteo Tonietto, Olof Strandberg, Niklas Mattsson-Carlgren, Sebastian Palmqvist, Michael J Pontecorvo, et al. Comparison of group-level and individualized brain regions for measuring change in l...

    2023

  11. [19]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. In ICLR (Poster) . OpenReview.net, 2018

    2018

  12. [20]

    Probabilistic disease progression modeling to characterize diagnostic uncertainty: application to staging and prediction in alzheimer's disease

    Marco Lorenzi, Maurizio Filippone, Giovanni B Frisoni, Daniel C Alexander, S \'e bastien Ourselin, Alzheimer's Disease Neuroimaging Initiative, et al. Probabilistic disease progression modeling to characterize diagnostic uncertainty: application to staging and prediction in al...

    2019

  13. [21]

    In vivo rate-determining steps of tau seed accumulation in Alzheimer's disease

    Georg Meisl, Eric Hidari, Kieren Allinson, Timothy Rittman, Sarah L Devos, Justin S Sanchez, Catherine K Xu, Karen E Duff, Keith A Johnson, James B Rowe, Bradley T Hyman, Tuomas P J Knowles, and David Klenerman. In vivo rate-determining steps of tau seed accumulation in Alzhei...

    2021

  14. [22]

    In vivo rate-determining steps of tau seed accumulation in alzheimer’s disease

    Georg Meisl, Eric Hidari, Kieren Allinson, Timothy Rittman, Sarah L DeVos, Justin S Sanchez, Catherine K Xu, Karen E Duff, Keith A Johnson, James B Rowe, et al. In vivo rate-determining steps of tau seed accumulation in alzheimer’s disease. Science advances, 7 0 (44): 0 eabh14...

    2021

  15. [23]

    Graph neural ordinary differential equations

    Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations. arXiv preprint arXiv:1911.07532, 2019

    1911 arXiv

  16. [24]

    A Network Diffusion Model of Disease Progression in Dementia

    Ashish Raj, Amy Kuceyeski, and Michael Weiner. A Network Diffusion Model of Disease Progression in Dementia . Neuron, 73 0 (6): 0 1204--1215, 3 2012 a . ISSN 08966273. doi:10.1016/j.neuron.2011.12.040

    2012 doi

  17. [25]

    A network diffusion model of disease progression in dementia

    Ashish Raj, Amy Kuceyeski, and Michael Weiner. A network diffusion model of disease progression in dementia. Neuron, 73 0 (6): 0 1204--1215, 2012 b

    2012

  18. [26]

    An Open MRI Dataset For Multiscale Neuroscience

    Jessica Royer et al. An Open MRI Dataset For Multiscale Neuroscience . Scientific Data, 9 0 (1), 12 2022. ISSN 20524463. doi:10.1038/s41597-022-01682-y

    2022 doi

  19. [27]

    Learning internal representations by error propagation, 1985

    David E Rumelhart, Geoffrey E Hinton, Ronald J Williams, et al. Learning internal representations by error propagation, 1985

    1985

  20. [28]

    Brain network communication: concepts, models and applications

    Caio Seguin, Olaf Sporns, and Andrew Zalesky. Brain network communication: concepts, models and applications. Nature reviews neuroscience, 24 0 (9): 0 557--574, 2023

    2023

  21. [29]

    Structured sequence modeling with graph convolutional recurrent networks

    Youngjoo Seo, Micha \" e l Defferrard, Pierre Vandergheynst, and Xavier Bresson. Structured sequence modeling with graph convolutional recurrent networks. In ICONIP (1) , volume 11301 of Lecture Notes in Computer Science, pages 362--373. Springer, 2018

    2018

  22. [30]

    Learning heat diffusion graphs

    Dorina Thanou, Xiaowen Dong, Daniel Kressner, and Pascal Frossard. Learning heat diffusion graphs. IEEE Trans. Signal Inf. Process. over Networks , 3 0 (3): 0 484--499, 2017

    2017

  23. [31]

    Graph attention networks

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li \` o , and Yoshua Bengio. Graph attention networks. CoRR, abs/1710.10903, 2017

    2017 arXiv

  24. [32]

    Multiphysics of Prionlike Diseases: Progression and Atrophy

    Johannes Weickenmeier, Ellen Kuhl, and Alain Goriely. Multiphysics of Prionlike Diseases: Progression and Atrophy . Physical Review Letters, 121 0 (15), 10 2018 a . ISSN 10797114. doi:10.1103/PhysRevLett.121.158101

    2018 doi

  25. [33]

    Multiphysics of prionlike diseases: Progression and atrophy

    Johannes Weickenmeier, Ellen Kuhl, and Alain Goriely. Multiphysics of prionlike diseases: Progression and atrophy. Physical review letters, 121 0 (15): 0 158101, 2018 b

    2018

  26. [34]

    Longitudinal predictive modeling of tau progression along the structural connectome

    Fan Yang, Samadrita Roy Chowdhury, Heidi IL Jacobs, Jorge Sepulcre, Van J Wedeen, Keith A Johnson, and Joyita Dutta. Longitudinal predictive modeling of tau progression along the structural connectome. NeuroImage, 237: 0 118126, 2021

    2021

  27. [35]

    Augmenting physical models with deep networks for complex dynamics forecasting

    Yuan Yin, Vincent Le Guen, J \'e r \'e mie Dona, Emmanuel de B \'e zenac, Ibrahim Ayed, Nicolas Thome, and Patrick Gallinari. Augmenting physical models with deep networks for complex dynamics forecasting. Journal of Statistical Mechanics: Theory and Experiment, 2021 0 (12): 0...

    2021

  28. [36]

    A data-driven model of biomarker changes in sporadic alzheimer's disease

    Alexandra L Young, Neil P Oxtoby, Pankaj Daga, David M Cash, Nick C Fox, Sebastien Ourselin, Jonathan M Schott, and Daniel C Alexander. A data-driven model of biomarker changes in sporadic alzheimer's disease. Brain, 137 0 (9): 0 2564--2577, 2014

    2014

  29. [37]

    Young, Neil P

    Alexandra L. Young, Neil P. Oxtoby, Sara Garbarino, Nick C. Fox, Frederik Barkhof, Jonathan M. Schott, and Daniel C. Alexander. Data-driven modelling of neurodegenerative disease progression: thinking outside the black box , 2024. ISSN 14710048

    2024

  30. [38]

    Spatio-temporal graph transformer networks for pedestrian trajectory prediction

    Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In ECCV (12) , volume 12357 of Lecture Notes in Computer Science, pages 507--523. Springer, 2020

    2020

  31. [39]

    Neural dynamics on complex networks

    Chengxi Zang and Fei Wang. Neural dynamics on complex networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 892--902, 2020

    2020

  32. [40]

    Gennatas, Joel H

    Juan Zhou, Efstathios D. Gennatas, Joel H. Kramer, Bruce L. Miller, and William W. Seeley. Predicting Regional Neurodegeneration from the Healthy Brain Functional Connectome . Neuron, 73 0 (6): 0 1216--1227, 3 2012 a . ISSN 08966273. doi:10.1016/j.neuron.2012.03.004

    2012 doi

  33. [41]

    Predicting regional neurodegeneration from the healthy brain functional connectome

    Juan Zhou, Efstathios D Gennatas, Joel H Kramer, Bruce L Miller, and William W Seeley. Predicting regional neurodegeneration from the healthy brain functional connectome. Neuron, 73 0 (6): 0 1216--1227, 2012 b

    2012

Pith tools

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