arxiv: 2604.06469 · v1 · submitted 2026-04-07 · 💻 cs.CV

Recognition: 1 theorem link

· Lean Theorem

Predicting Alzheimer's disease progression using rs-fMRI and a history-aware graph neural network

Austin Toma, Connor Laming, Huirong Fu, Mahdi Moghaddami, Mohammad-Reza Siadat

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:38 UTC · model grok-4.3

classification 💻 cs.CV

keywords Alzheimer's diseasers-fMRIgraph neural networkrecurrent neural networkdisease progressioncognitive impairmentfunctional connectivitystage transition prediction

0 comments

The pith

A history-aware graph neural network predicts Alzheimer's stage transitions from rs-fMRI at 82.9% accuracy.

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

The paper builds a model that takes resting-state fMRI scans, converts them into functional connectivity graphs, and feeds the full sequence of a subject's visits into a graph neural network augmented with a recurrent block. This setup lets the model account for prior visits and the time gaps between them while predicting whether the next visit will show cognitively normal, mild cognitive impairment, or Alzheimer's disease. A sympathetic reader would care because the model reaches 82.9% overall accuracy and 68.8% on the especially difficult CN-to-MCI shift, raising the possibility of earlier interventions that slow progression. The approach also tolerates irregular visit spacing and missing scans, which matches real clinical records.

Core claim

The authors claim that a GNN model equipped with an RNN block can process an entire history of rs-fMRI-derived functional connectivity graphs, incorporate visit-distance features, and output the next cognitive stage for each subject, achieving 82.9% accuracy across CN, MCI, and AD transitions on 303 subjects and 68.8% accuracy specifically on CN-to-MCI conversions, while remaining robust when some visits are absent.

What carries the argument

A graph neural network whose message-passing layers receive functional connectivity graphs from each rs-fMRI visit and whose output is sequenced by an RNN block that also receives explicit visit-interval values, allowing the network to model longitudinal progression.

If this is right

Timely prediction of CN-to-MCI transitions could allow interventions before full AD develops.
The model continues to function when visits are irregularly spaced or partially missing, matching typical clinical data.
rs-fMRI functional connectivity alone supplies usable information for forecasting MCI or AD onset.
The same architecture could be combined with other modalities to raise accuracy further.

Where Pith is reading between the lines

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

If the same graph-plus-history structure works on larger or multi-site data, the method could support population-level screening programs.
The robustness to missing visits opens the possibility of applying the model to retrospective medical records that lack complete follow-up.
Architectures that fuse GNNs on brain graphs with RNNs on time sequences may transfer to progression modeling in other neurodegenerative conditions.

Load-bearing premise

The functional connectivity graphs extracted from rs-fMRI plus the sequence of visit times contain enough generalizable signal to forecast future stage changes without being tied to the particular 303-subject cohort or requiring dataset-specific fixes.

What would settle it

Train the model on the reported 303-subject set and then test it unchanged on an independent external cohort that supplies rs-fMRI scans, visit dates, and confirmed stage labels at the subsequent visit; accuracy below 70% on the full task or below 60% on CN-to-MCI would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.06469 by Austin Toma, Connor Laming, Huirong Fu, Mahdi Moghaddami, Mohammad-Reza Siadat.

**Figure 1.** Figure 1: Distribution of subjects based on their number of available rs-fMRI scans. The blue bars represent stable [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Distribution of subjects based on their diagnosis in their last visit [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of our workflow, from raw rs-fMRI scans to the binary prediction [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Architecture of our HA-GNN model [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: Architecture of the GCN pretraining model [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

read the original abstract

Alzheimer's disease (AD) is a neurodegenerative disorder that affects more than seven million people in the United States alone. AD currently has no cure, but there are ways to potentially slow its progression if caught early enough. In this study, we propose a graph neural network (GNN)-based model for predicting whether a subject will transition to a more severe stage of cognitive impairment at their next clinical visit. We consider three stages of cognitive impairment in order of severity: cognitively normal (CN), mild cognitive impairment (MCI), and AD. We use functional connectivity graphs derived from resting-state functional magnetic resonance imaging (rs-fMRI) scans of 303 subjects, each with a different number of visits. Our GNN-based model incorporates a recurrent neural network (RNN) block, enabling it to process data from the subject's entire visit history. It can also work with irregular time gaps between visits by incorporating visit distance information into our input features. Our model demonstrates robust predictive performance, even with missing visits in the subjects' visit histories. It achieves an accuracy of 82.9%, with an especially impressive accuracy of 68.8% on CN to MCI conversions - a task that poses a substantial challenge in the field. Our results highlight the effectiveness of rs-fMRI in predicting the onset of MCI or AD and, in conjunction with other modalities, could offer a viable method for enabling timely interventions to slow the progression of cognitive impairment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adds an RNN block to a GNN to handle full visit history and irregular gaps in rs-fMRI graphs for AD stage prediction, but the 82.9% accuracy (68.8% on CN-to-MCI) rests on a single 303-subject cohort without clear validation details.

read the letter

The core contribution is a GNN that feeds functional connectivity graphs into an RNN so the model can use a subject's complete sequence of visits, including time gaps between them. This lets it work with missing scans and irregular timing, which is a practical step for real longitudinal rs-fMRI data. The reported numbers are 82.9% overall accuracy and 68.8% on the CN-to-MCI transitions that matter most for early intervention. That architecture choice is a straightforward extension of existing GNN and RNN ideas rather than a big conceptual leap, but it fits the problem of variable-length histories reasonably well. The 303-subject set is typical for this domain and the claim that it tolerates missing visits is useful if it holds. The main weakness is the missing validation information. The abstract gives accuracies but says nothing about subject-independent splits, cross-validation folds, class imbalance handling, or baseline comparisons. With a modest cohort and relatively rare conversion events, those numbers could easily come from cohort-specific patterns or leakage instead of robust signals. If the full paper shows proper held-out testing and statistical checks, the result strengthens; right now it is hard to judge. This work is aimed at researchers combining graph models with longitudinal neuroimaging for neurodegenerative prediction. A reader already working on multimodal AD biomarkers or time-aware GNNs could extract the RNN integration idea and test it on their own data. It is worth sending to peer review so referees can examine the exact splits, architecture details, and any ablation results; the idea is coherent enough that a careful review would clarify whether the performance generalizes.

Referee Report

2 major / 1 minor

Summary. The paper proposes a GNN model augmented with an RNN block to predict transitions between cognitive stages (CN, MCI, AD) from rs-fMRI functional connectivity graphs. It processes longitudinal visit histories with irregular intervals and missing data for a cohort of 303 subjects, reporting 82.9% overall accuracy and 68.8% accuracy specifically on CN-to-MCI conversions.

Significance. If the performance holds under proper validation, the history-aware architecture offers a practical way to leverage rs-fMRI for early detection of progression, addressing a clinically relevant gap where CN-to-MCI prediction is difficult. The explicit handling of visit timing and missing data is a constructive modeling choice for real-world longitudinal cohorts.

major comments (2)

Abstract: The reported accuracies (82.9% overall, 68.8% on CN-to-MCI) are presented without any information on validation splits, subject-independent cross-validation, class balance, statistical tests, or handling of missing visits. With only 303 subjects and infrequent conversion events, this absence directly undermines assessment of whether the numbers reflect generalizable signals or cohort-specific overfitting.
Methods/Results (inferred from abstract claims): The central assertion that rs-fMRI graphs plus RNN-processed history suffice for robust prediction lacks supporting details on baseline comparisons, ablation of the RNN or visit-distance features, or external test-set performance. Without these, the 68.8% CN-to-MCI figure cannot be evaluated as evidence of predictive utility rather than leakage or post-hoc tuning.

minor comments (1)

Abstract: Provide the number of subjects per diagnostic category and the observed conversion rates to allow readers to interpret the reported accuracies in context.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which identify key areas where additional details will strengthen the manuscript's clarity and rigor. We address each major comment below and will incorporate the suggested revisions.

read point-by-point responses

Referee: Abstract: The reported accuracies (82.9% overall, 68.8% on CN-to-MCI) are presented without any information on validation splits, subject-independent cross-validation, class balance, statistical tests, or handling of missing visits. With only 303 subjects and infrequent conversion events, this absence directly undermines assessment of whether the numbers reflect generalizable signals or cohort-specific overfitting.

Authors: We agree that the abstract and main text require more explicit details on the validation protocol to support the reported performance. In the revised manuscript, we will expand the abstract and Methods section to describe the subject-independent cross-validation procedure, how class imbalance was addressed (via weighted loss functions given the rarity of conversions), the statistical tests applied (e.g., McNemar's test or bootstrap confidence intervals), and the specific mechanisms for handling missing visits and irregular intervals. We will also report cohort statistics on conversion event frequency and discuss potential overfitting risks in light of the sample size of 303 subjects, ensuring readers can better evaluate generalizability. revision: yes
Referee: Methods/Results (inferred from abstract claims): The central assertion that rs-fMRI graphs plus RNN-processed history suffice for robust prediction lacks supporting details on baseline comparisons, ablation of the RNN or visit-distance features, or external test-set performance. Without these, the 68.8% CN-to-MCI figure cannot be evaluated as evidence of predictive utility rather than leakage or post-hoc tuning.

Authors: We acknowledge that baseline comparisons and ablations are essential to isolate the contributions of the history-aware components. In the revised manuscript, we will add a dedicated ablation study section comparing the full model against (i) a standard GNN without the RNN block, (ii) a version without visit-distance features, and (iii) non-graph baselines such as SVMs or LSTMs on vectorized features. We will also clarify the cross-validation design (subject-independent folds to avoid leakage) and discuss the absence of a fully held-out external test set due to the single-cohort nature of the data, while providing internal validation metrics with confidence intervals. These additions will allow direct assessment of whether the 68.8% CN-to-MCI accuracy reflects genuine predictive utility. revision: yes

Circularity Check

0 steps flagged

Empirical ML evaluation on cohort data shows no circularity

full rationale

The paper presents a GNN model augmented with an RNN block to process subject visit histories from rs-fMRI functional connectivity graphs, reporting empirical accuracies of 82.9% overall and 68.8% on CN-to-MCI transitions after training and evaluation on the 303-subject dataset. No mathematical derivations, equations, uniqueness theorems, or self-citations are invoked that reduce the reported performance metrics to quantities defined by construction from fitted parameters or prior author results. The central claims rest on standard supervised learning outcomes rather than any self-referential reduction, making the work self-contained as an empirical study.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The performance claim rests on standard neural network training assumptions plus the domain premise that rs-fMRI graphs encode progression signals; no independent evidence or formal verification is supplied in the abstract.

free parameters (1)

GNN and RNN architecture hyperparameters
Layer sizes, learning rates, and other tuning choices that determine the reported accuracies but are not disclosed.

axioms (2)

domain assumption Functional connectivity graphs derived from rs-fMRI contain information predictive of future cognitive stage transitions
Directly invoked by using these graphs as the primary input to the model.
domain assumption Incorporating visit history via RNN improves prediction over single-scan models
The model design and reported gains on CN-to-MCI assume this holds.

pith-pipeline@v0.9.0 · 5573 in / 1200 out tokens · 62655 ms · 2026-05-10T18:38:54.199374+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
Our GNN-based model incorporates a recurrent neural network (RNN) block... functional connectivity graphs derived from rs-fMRI... 303 subjects... accuracy of 82.9%

Reference graph

Works this paper leans on

21 extracted references · 1 canonical work pages

[1]

Alzheimer’s disease,

Scheltens, P., Blennow, K., Breteler, M. M. B., de Strooper, B., Frisoni, G. B., Salloway, S., and Van der Flier, W. M., “Alzheimer’s disease,”The Lancet388(10043), 505–517 (2016)

2016
[2]

Dementia prevention, intervention, and care: 2020 report of the lancet commission,

Livingston, G., Huntley, J., Sommerlad, A., Ames, D., Ballard, C., Banerjee, S., Brayne, C., Burns, A., Cohen-Mansfield, J., Cooper, C., Costafreda, S. G., Dias, A., Fox, N., Gitlin, L. N., Howard, R., Kales, H. C., Kivim¨ aki, M., Larson, E. B., Ogunniyi, A., Orgeta, V., Ritchie, K., Rockwood, K., Sampson, E. L., Samus, Q., Schneider, L. S., Selbæk, G., ...

2020
[3]

The alzheimer’s disease neuroimaging initiative (adni): Mri methods,

Jack, C. R., Bernstein, M. A., Fox, N. C., Thompson, P. M., Alexander, G. E., Harvey, D. J., Borowski, B. J., Britson, P. J., Whitwell, J. L., Ward, C. P., Dale, A. M., Felmlee, J. P., Gunter, J. L., Hill, D. L. G., Killiany, R. J., Schuff, N., Fox-Bosetti, S., Lin, C., Studholme, C., DeCarli, C., Krueger, G., Ward, H. A., Metzger, G. J., Scott, K. T., Ma...

2008
[4]

The role of neuroimag- ing in alzheimer’s disease: implications for the diagnosis, monitoring disease progression, and treatment,

Mulumba, J., Duan, R., Luo, B., Wu, J., Sulaiman, M., Wang, F., and Yang, Y., “The role of neuroimag- ing in alzheimer’s disease: implications for the diagnosis, monitoring disease progression, and treatment,” Exploration of Neuroscience4(Feb. 2025)

2025
[5]

Brain inflammation, degeneration, and plasticity in multiple sclerosis,

Granziera, C. and Sprenger, T., “Brain inflammation, degeneration, and plasticity in multiple sclerosis,” in [Brain Mapping], Toga, A. W., ed., 917–927, Academic Press, Waltham (2015)

2015
[6]

Inductive Representation Learning on Large Graphs

Hamilton, W. L., Ying, Z., and Leskovec, J., “Inductive representation learning on large graphs,” ArXivabs/1706.02216(2017)

work page Pith review arXiv 2017
[7]

Braingnn: Interpretable brain graph neural network for fmri analysis,

Li, X., Zhou, Y., Gao, S., Dvornek, N. C., Zhang, M., Zhuang, J., Gu, S., Scheinost, D., Staib, L. H., Ventola, P., and Duncan, J. S., “Braingnn: Interpretable brain graph neural network for fmri analysis,” bioRxiv(2020)

2020
[8]

Transformer model for alzheimer’s disease progression prediction using longitudinal visit sequences,

Moghaddami, M., Schubring, C., and Siadat, M., “Transformer model for alzheimer’s disease progression prediction using longitudinal visit sequences,” in [Proceedings of the sixth Conference on Health, Inference, and Learning], Xu, X. O., Choi, E., Singhal, P., Gerych, W., Tang, S., Agrawal, M., Subbaswamy, A., Sizikova, E., Dunn, J., Daneshjou, R., Sarker...

2025
[9]

Long short-term memory,

Hochreiter, S. and Schmidhuber, J., “Long short-term memory,”Neural computation9(8), 1735–1780 (1997)

1997
[10]

Learning phrase representations using rnn encoder–decoder for statistical machine translation,

Cho, K., van Merrienboer, B., C ¸ aglar G¨ ul¸ cehre, Bahdanau, D., Bougares, F., Schwenk, H., and Bengio, Y., “Learning phrase representations using rnn encoder–decoder for statistical machine translation,” in [Conference on Empirical Methods in Natural Language Processing], (2014)

2014
[11]

Tadpole challenge: Prediction of longitudinal evolution in alzheimer’s disease.,

Marinescu, R. V., Oxtoby, N. P., Young, A. L., Bron, E., Toga, A. W., Weiner, M. W., Barkhof, F., Fox, N. C., Klein, S., Alexander, D. C., the EuroPOND Consortium, and the Alzheimer’s Disease Neuroimag- ing Initiative, F., “Tadpole challenge: Prediction of longitudinal evolution in alzheimer’s disease.,”arXiv: Populations and Evolution(2018)

2018
[12]

fmriprep: a robust preprocessing pipeline for functional mri,

Esteban, O., Markiewicz, C. J., Blair, R. W., Moodie, C. A., Isik, A. I., Erramuzpe, A., Kent, J. D., Goncalves, M., DuPre, E., Snyder, M., Oya, H., Ghosh, S. S., Wright, J., Durnez, J., Poldrack, R. A., and Gorgolewski, K. J., “fmriprep: a robust preprocessing pipeline for functional mri,”Nature Methods16, 111–116 (Dec. 2018)

2018
[13]

The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments,

Gorgolewski, K. J., Auer, T., Calhoun, V. D., Craddock, R. C., Das, S., Duff, E. P., Flandin, G., Ghosh, S. S., Glatard, T., Halchenko, Y. O., Handwerker, D. A., Hanke, M., Keator, D. B., Li, X., Michael, Z., Maumet, C., Nichols, B. N., Nichols, T. E., Pellman, J., Poline, J.-B., Rokem, A. S., Schaefer, G., Sochat, V. V., Triplett, W., Turner, J. A., Varo...

2016
[14]

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

Schaefer, A., Kong, R., Gordon, E. M., Laumann, T. O., Zuo, X.-N., Holmes, A. J., Eickhoff, S. B., and Yeo, B. T. T., “Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity mri,”Cerebral Cortex28, 3095–3114 (July 2017)

2017
[15]

Graphnorm: A principled approach to accelerating graph neural network training,

Cai, T., Luo, S., Xu, K., He, D., Liu, T.-Y., and Wang, L., “Graphnorm: A principled approach to accelerating graph neural network training,” in [International Conference on Machine Learning], (2020)

2020
[16]

Practical bayesian optimization of machine learning algo- rithms,

Snoek, J., Larochelle, H., and Adams, R. P., “Practical bayesian optimization of machine learning algo- rithms,” in [Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2],NIPS’12, 2951–2959, Curran Associates Inc., Red Hook, NY, USA (2012)

2012
[17]

Focal loss for dense object detection,

Lin, T.-Y., Goyal, P., Girshick, R. B., He, K., and Doll´ ar, P., “Focal loss for dense object detection,”2017 IEEE International Conference on Computer Vision (ICCV), 2999–3007 (2017)

2017
[18]

The use of the area under the roc curve in the evaluation of machine learning algorithms,

Bradley, A. P., “The use of the area under the roc curve in the evaluation of machine learning algorithms,” Pattern Recognit.30, 1145–1159 (1997)

1997
[19]

Interpretable temporal graph neural network for prognostic prediction of alzheimer’s disease using longitudinal neuroimaging data,

Kim, M., Kim, J., Qu, J., Huang, H., Long, Q., Sohn, K.-A., Kim, D., and Shen, L., “Interpretable temporal graph neural network for prognostic prediction of alzheimer’s disease using longitudinal neuroimaging data,” in [2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)], 1381–1384, IEEE (Dec. 2021)

2021
[20]

Predicting the conversion from mild cognitive impairment to alzheimer’s disease using an explainable ai approach,

Grammenos, G., Vrahatis, A. G., Vlamos, P., Palejev, D., and Exarchos, T., “Predicting the conversion from mild cognitive impairment to alzheimer’s disease using an explainable ai approach,”Information15, 249 (Apr. 2024)

2024
[21]

A self-guided multimodal approach to enhancing graph representation learning for alzheimer’s diseases,

Wang, Z., Bao, R., Wu, Y., Liu, G., Yang, L., Zhan, L., Zheng, F., Jiang, W., and Zhang, Y., “A self-guided multimodal approach to enhancing graph representation learning for alzheimer’s diseases,” (2024)

2024