Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

NeuroMoE: A Transformer-Based Mixture-of-Experts Framework for Multi-Modal Neurological Disorder 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 A gated mixture-of-experts that fuses three MRI modalities with clinical and serum data can classify PD, iRBD, and healthy controls at 82.47 percent accuracy.

desk verdict The reported 82.47% accuracy likely re-describes the diagnostic labels baked into the clinical inputs (PSG, H&Y, disease duration), so the paper's central claim about multimodal MoE fusion is unsupported by the current evaluation. read the letter →

arxiv 2506.14970 v1 pith:M3O75H67 submitted 2025-06-17 eess.IV cs.CVcs.LG

Wajih Hassan Raza , Aamir Bader Shah , Yu Wen , Yidan Shen , Juan Diego Martinez Lemus , Mya Caryn Schiess , Timothy Michael Ellmore , Renjie Hu
show 1 more author
Xin Fu
This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords mixture-of-expertstransformermultimodalMRIfusionneurologicaldisorderclassificationParkinson'sdiseaseidiopathicREMsleepbehaviorgatingnetworkclinicalbiomarkers
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 a transformer-based mixture-of-experts (MoE) architecture can fuse anatomical, diffusion, and functional MRI with serum and clinical measurements to classify Parkinson's disease, idiopathic REM sleep behavior disorder, and healthy controls at 82.47 percent validation accuracy, beating single-modality baselines by more than ten points. If true, it would mean that adaptive, patient-specific weighting of modality experts, rather than any single scan type, is the key to distinguishing early and overlapping neurodegenerative states. The authors intend NeuroMoE as a clinically usable diagnostic aid for prodromal and established disease, with lightweight experts that could run in hospital systems. The design centers on modality-specific transformer encoders feeding four experts that a clinical-feature-driven gating network combines per patient.

What carries the argument

The load-bearing component is the Personalized Adaptive MoE block: four modality-specific experts (functional, anatomical, diffusion, and serum/clinical), each a small feed-forward network producing a class prediction, combined by a gating network that maps clinical features through two hidden layers into a softmax weight vector $\omega \in \mathbb{R}^4$. The final prediction is the weighted sum $P_{\mathrm{out}} = \sum_{i=1}^4 \omega_i P_i$, with a balance regularizer that penalizes gating weights drifting from a uniform average. This mechanism does the work of deciding, per patient, how much each modality contributes to the diagnosis.

What would settle it

Retrain the model on the same cohort while removing the polysomnography result, Hoehn and Yahr stage, and disease duration from the clinical inputs, then compare accuracy against the full model: if accuracy stays near 82.47 percent, the multimodal architecture carries the signal, and if it falls toward the roughly 70 percent baseline range, the gain largely comes from label-encoding clinical variables rather than from MRI fusion.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a gated mixture-of-experts over modality-specific transformer encoders makes multi-modal neurological disorder classification work on a real clinical cohort where single-modality models plateau below 60 percent accuracy. NeuroMoE reaches 82.47 percent accuracy (F1 81.25 percent) on the three-way PD, iRBD, and healthy-control task, with ablations showing that removing any modality drops performance and that removing the gating mechanism drops accuracy most sharply, to 65.17 percent. The paper interprets this as evidence that complementary modalities carry distinct disease information and that dynamically weighting them by clinical context is what lets the model separate overlapping disease states.

Load-bearing premise

The load-bearing premise is that the clinical variables fed into the model, above all the sleep-study result that defines one of the three diagnostic groups and the staging scale and illness duration that define the other, are neutral predictors rather than pieces of the diagnostic gold standard.

Editorial extensions

If this is right

  • A single model can separate PD, iRBD, and healthy controls at 82.47 percent accuracy on a real clinical cohort, more than ten points above the best single-modality baseline, so multimodal fusion is the source of the reported gain.
  • Because the gating network weights each patient's experts based on clinical features, the model can in principle show clinicians which modality drove a given prediction, adding interpretability to the diagnosis.
  • The ablation results imply that every input modality matters: removing the serum/clinical encoder costs the most among modality removals, and removing the gate costs more than any single modality removal, so both breadth of modalities and adaptive weighting are needed.
  • The framework is designed to extend beyond PD and iRBD to other synucleinopathies and to longitudinal prediction of iRBD-to-PD conversion, once corresponding datasets are available.

Reading between the lines

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

  • One consequence the authors do not test: the clinical variables used to drive the gate, particularly the sleep-study result that defines iRBD and the staging scale and illness duration that define PD, may partly encode the diagnostic labels, so the reported accuracy should be re-measured with those variables withheld.
  • A natural extension would be to gate on imaging-derived or other non-label variables; if accuracy held, it would show that the adaptive fusion itself, rather than diagnostic content in the gating inputs, is responsible for the improvement.
  • If the result reproduces on an external dataset, the per-patient gating weights could serve as a map of which modality is most informative at different disease stages, something the current average-utilization figure does not reveal.
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. NeuroMoE (arXiv:2506.14970) proposes a transformer-based Mixture-of-Experts classifier that fuses aMRI, DTI, fMRI, clinical assessments, and serum biomarkers from a proprietary 113-subject cohort (41 PD, 44 iRBD, 28 HC). The paper reports 82.47% test accuracy and >10% improvement over single-modality baselines, and argues that adaptive gating of modality-specific experts is responsible for the gain.

Significance. The clinical question—distinguishing PD, iRBD, and HC using multimodal data—is important, and the proposed architecture (modality-specific encoders, gated MoE, balance regularization) is a plausible design. However, the evaluation does not support the stated claim. The clinical features include the diagnostic gold standard itself (PSG, H&Y, disease duration), so the reported accuracy may largely re-describe the labels. The comparison is also unfair because the baselines receive no clinical features. If the leakage were removed, the central contribution would need to be re-established. The paper does include an ablation study and expert-utilization analysis, but those results are consistent with reliance on the leaked features rather than on multimodal MRI fusion.

major comments (4)
  1. [Section III, Table I] The clinical feature vector contains PSG result, Hoehn & Yahr (H&Y) stage, and disease duration. PSG is the polysomnographic gold standard for iRBD (42 of 44 iRBD subjects are PSG=1 versus 0 of 28 HC), H&Y is a Parkinson's disease severity scale that is 1.73±0.72 for PD and 0.00 for both iRBD and HC, and disease duration is 0.00 for every HC. A simple decision rule on these three fields can already separate the classes almost perfectly. Since the gating network in Eq. (2) is itself conditioned on clinical features (Section IV.D), the 82.47% accuracy in Table II does not demonstrate that the multimodal MoE architecture learns useful imaging representations; it likely re-describes the diagnostic labels. This invalidates the central claim of >10% improvement over single-modality baselines, which never see these clinical fields.
  2. [Section V.A] The best model is selected on the test set: the text states that 'the best model was saved based on test accuracy,' and there is only a single 80/20 split. With N=113, the test partition contains approximately 23 subjects, so the reported 82.47% figure has large sampling uncertainty. No confidence intervals, bootstrap estimates, or significance tests are reported. The >10% gap over baselines is therefore not statistically substantiated.
  3. [Section V.B, Table III, Fig. 4] The ablation study shows the largest accuracy drop when the serum/clinical encoder is removed (82.47% to 75.12%), and Fig. 4 shows the serum/clinical expert has the highest average utilization. This pattern is exactly what would be expected if the model relies on label-overlapping clinical variables; it does not support the claim that imaging modalities contribute complementary information. The authors should report results after removing PSG, H&Y, and disease duration, and should include a clinical-variables-only baseline or a multimodal baseline that receives the same clinical variables.
  4. [Table II] The baselines are exclusively single-modality MRI models. Because the full NeuroMoE receives clinical and serum data in addition to MRI, the comparison conflates modality fusion with the addition of extra information sources. A fair comparison would include a multimodal baseline that also receives the same clinical variables, so that the effect of the MoE gating mechanism can be isolated from the effect of having access to the clinical gold-standard variables.
minor comments (6)
  1. [Section IV.B] In the serum/clinical preprocessing description, 'P&G results' should read 'PSG results.'
  2. [Eq. (1)] The notation is unclear: P E is used both as a function and as an operand, and M P is not defined. Please define all symbols or provide a clearer equation.
  3. [Abstract and Section V.A] The abstract says 'validation accuracy,' while Section V.A describes a train/test split; use consistent terminology to clarify whether the reported number is a validation or test accuracy.
  4. [Section V.A] The sentence 'Since our dataset is unique and one of a kind...' is informal and does not justify the absence of external validation or cross-validation; please rephrase and provide a concrete rationale.
  5. [Table II] No confidence intervals or p-values are reported for any comparison; add them or justify their absence, especially given the small test set.
  6. [References [34], [35]] The Dynamic Image and LiuNet baselines were originally developed for Alzheimer's disease classification; the authors should explain why these architectures are appropriate and competitive for PD/iRBD classification.

Circularity Check

1 steps flagged · score 8.0 of 10

Clinical inputs re-describe the diagnostic labels: PSG, H&Y, and disease duration make the 82.47% accuracy a re-reading of the gold standard rather than a validation of multimodal MoE fusion.

  1. self definitional [Section III (Dataset), Table I; Section IV.D (Gating); Table III (Ablation)]
    "Section III: 'had polysomnographically confirmed iRBD (44 participants)'; Table I: 'PSG result 0 (18), 1 (23) 0 (2), 1 (42) 0 (28), 1 (0)', 'H&Y 1.73 ± 0.72 0.00 ± 0.00 0.00 ± 0.00', 'Disease Duration 6.11 ± 3.41 6.45 ± 3.12 0.00 ± 0.00'."

    The target classes are defined by these very inputs: iRBD is polysomnographically confirmed, so the PSG result is the gold-standard diagnostic label; H&Y is a PD severity scale that is 0.00 for all non-PD subjects; disease duration is 0.00 for every HC. A threshold rule on these three clinical fields can separate PD, iRBD, and HC almost perfectly without using any MRI or learned MoE representation. The reported 82.47% accuracy and the >10% margin over single-MRI baselines therefore do not validate the multimodal architecture; they are substantially forced by input-label overlap. The gating network takes clinical features as input (Eq. 2), and Fig.

full rationale

The load-bearing claim is that NeuroMoE's transformer-based mixture-of-experts multimodal fusion achieves 82.47% accuracy and outperforms baselines by over 10% (Abstract; Section V.B). This claim is undermined by construction because the serum/clinical feature set includes variables that are definitional of the outcome classes. Section III states that iRBD participants were 'polysomnographically confirmed,' and Table I shows PSG result alone is 1 for 42/44 iRBD and 0/28 HC, nearly separating those classes; H&Y is 1.73 for PD but 0.00 for both iRBD and HC; disease duration is 0.00 for all HC. A trivial linear or threshold classifier on these fields can reproduce or exceed the reported accuracy without MRI or the MoE machinery. Since none of the Table II baselines receive clinical features, the >10% improvement is not evidence for the architecture but evidence that the test inputs encode the gold standard. Fig. 4 and Table III further show the model relies most on the serum/clinical expert. Additionally, Section V.A states 'the best model was saved based on test accuracy,' so the reported test accuracy is a selected maximum rather than an unbiased estimate; this is an evaluation leak, though not the core circularity. The NeuroMoE architecture itself is not circular, but the central evaluation claim reduces to input-label overlap, warranting a score of 8.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central performance claim rests on several unstated domain assumptions, most critically that diagnostic-defining clinical variables (PSG, H&Y, disease duration) can be used as predictive inputs without leaking the labels. It also assumes the pre-processing and small-sample training setup preserve generalizable signal. The listed free parameters are hyperparameters chosen empirically, none of which invalidates the method by themselves, but the clinical-feature assumption is the key ledger item.

free parameters (6)
  • Number of experts = 4
    Set to match the four modality encoders; chosen by design, not derived.
  • Patch size = 4x4x4 voxels
    Chosen as a hyperparameter for volumetric patch embedding; no justification given.
  • Transformer depth K = unspecified
    Paper says attention heads and transformer depth were chosen based on empirical results, but exact values are not reported.
  • Learning rate = 0.001
    Initial learning rate for Adam optimizer, chosen by convention/empirically.
  • Dropout rate = 30%
    Applied between layers of serum/clinical encoder and gating network; chosen by hand.
  • Gating regularization coefficient = not reported
    The balance regularization term R_balance is added but its weight is never specified.
assumptions (5)
  • domain assumption PSG result, H&Y stage, and disease duration are valid predictive covariates and not components of the diagnostic labels.
    The model uses these as inputs to classify PD/iRBD/HC; if they encode the gold standard, the reported accuracy is inflated. This assumption is unflagged in the paper.
  • domain assumption The FSL pre-processing pipeline (BET, FLIRT/FNIRT registration to MNI, intensity normalization) preserves disease-relevant signal across subjects and modalities.
    Section IV.B describes the pipeline without validation that it does not introduce artifacts or remove discriminative information.
  • domain assumption Mean imputation for 5% missing clinical/serum values is unbiased and does not distort class distributions.
    Section IV.B states mean imputation; no sensitivity analysis is provided.
  • domain assumption Transformer encoders with 4x4x4 patch embeddings can learn meaningful spatial representations from only 90 training subjects.
    The paper trains a transformer on a very small dataset without pre-training or data augmentation; the ability to generalize is assumed.
  • domain assumption The single 80/20 split yields a test set representative of the population.
    With N=113, the test set has ~23 samples; the paper does not stratify or cross-validate, and class imbalance may bias the estimate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeuroMoE: A Transformer-Based Mixture-of-Experts Framework for Multi-Modal Neurological Disorder Classification." pith.science (2026). https://pith.science/paper/M3O75H67

@misc{pith2026250614970,
  author       = {Pith},
  title        = {Pith review of: NeuroMoE: A Transformer-Based Mixture-of-Experts Framework for Multi-Modal Neurological Disorder Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3O75H67}},
  note         = {Machine review of arXiv:2506.14970}
}
read the original abstract

The integration of multi-modal Magnetic Resonance Imaging (MRI) and clinical data holds great promise for enhancing the diagnosis of neurological disorders (NDs) in real-world clinical settings. Deep Learning (DL) has recently emerged as a powerful tool for extracting meaningful patterns from medical data to aid in diagnosis. However, existing DL approaches struggle to effectively leverage multi-modal MRI and clinical data, leading to suboptimal performance. To address this challenge, we utilize a unique, proprietary multi-modal clinical dataset curated for ND research. Based on this dataset, we propose a novel transformer-based Mixture-of-Experts (MoE) framework for ND classification, leveraging multiple MRI modalities-anatomical (aMRI), Diffusion Tensor Imaging (DTI), and functional (fMRI)-alongside clinical assessments. Our framework employs transformer encoders to capture spatial relationships within volumetric MRI data while utilizing modality-specific experts for targeted feature extraction. A gating mechanism with adaptive fusion dynamically integrates expert outputs, ensuring optimal predictive performance. Comprehensive experiments and comparisons with multiple baselines demonstrate that our multi-modal approach significantly enhances diagnostic accuracy, particularly in distinguishing overlapping disease states. Our framework achieves a validation accuracy of 82.47\%, outperforming baseline methods by over 10\%, highlighting its potential to improve ND diagnosis by applying multi-modal learning to real-world clinical data.

Figures

Figures reproduced from arXiv: 2506.14970 by the authors.

Figure 1
Figure 1. Overall architecture of NeuroMoE. • Clinical Biomarkers includes several standardized scoring scales to assess disease severity and symptoms: the Modified Hoehn & Yahr Scale (H&Y) for disease stage, the Unified Parkinson’s Disease Rating Scale (UPDRS) Part III for motor function, the Timed Up and Go (TUG) Test for gait function, the Montreal Cognitive Assessment (MoCA) and the University of Pennsylva￾nia Smell Ident… view at source ↗
Figure 2
Figure 2. Pre-Processing Pipeline for Multi-Modal MRI. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 1
Figure 1. The input passes through three full-connected layers. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figures from the paper (2 more)
Figure 3
Figure 3. Figure 3: Workflow of expert utilization via dynamic gating. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]
Figure 4
Figure 4. Figure 4: Average expert utilization with regularization. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Lightweight Chunk Selection for Mobile Retrieval-Augmented Generation

    cs.LG 2026-08 conditional novelty 5.0 of 10

    Combining question hidden states, MoE routing signals, and chunk embeddings in a compact MLP improves rank-1 evidence chunk selection for mobile RAG by 2.49 points on average across TriviaQA, PopQA, and MS MARCO.

Reference graph

Works this paper leans on

35 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    The emerging evidence of the parkinson pandemic,

    E. R. Dorsey, T. Sherer, M. S. Okun, B. R. Bloem, P. Brundin, J. W. Langston, and B. R. Bloem, “The emerging evidence of the parkinson pandemic,”Journal of Parkinson’s Disease, vol. 8, no. s1, pp. S3–S8, 2018, pMID: 30584159

    work page 2018

  2. [2]

    Mds clinical diagnostic criteria for parkinson’s disease,

    R. B. Postuma, D. Berg, M. Stern, W. Poewe, C. W. Olanow, W. Oertel, J. Obeso, K. Marek, I. Litvan, A. E. Lang, G. Halliday, C. G. Goetz, T. Gasser, B. Dubois, P. Chan, B. R. Bloem, C. H. Adler, and G. Deuschl, “Mds clinical diagnostic criteria for parkinson’s disease,” Movement Disorders, vol. 30, no. 12, pp. 1591–1601, 2015

    work page 2015

  3. [3]

    Preclinical biomarkers of parkinson disease,

    Y . Wu, W. Le, and J. Jankovic, “Preclinical biomarkers of parkinson disease,”Archives of Neurology, vol. 68, no. 1, pp. 22–30, 01 2011

    work page 2011

  4. [4]

    Clinical utility of synuclein skin biopsy in the diagnosis and evaluation of synucle- inopathies,

    J. R. Isaacson, R. Freeman, and C. H. Gibbons, “Clinical utility of synuclein skin biopsy in the diagnosis and evaluation of synucle- inopathies,”Front. Neurol., vol. 15, p. 1510796, Dec. 2024

    work page 2024

  5. [5]

    Factors associated with phenoconversion of idiopathic rapid eye movement sleep behavior disorder: a prospective study,

    Y . Yuan, Y . Li, H. Zhang, Y . Zang, X. Liu, Y . Hou, S. Zhan, Y . Cai, W. Mao, and P. Chan, “Factors associated with phenoconversion of idiopathic rapid eye movement sleep behavior disorder: a prospective study,”npj Parkinson’s Disease, vol. 11, no. 1, p. 10, Jan 2025

    work page 2025

  6. [6]

    Evolution patterns of probable rem sleep behavior disorder predicts parkinson’s disease progression,

    G. Ye, X. Xu, L. Zhou, A. Zhao, L. Zhu, and J. Liu, “Evolution patterns of probable rem sleep behavior disorder predicts parkinson’s disease progression,”npj Parkinson’s Disease, vol. 8, no. 1, p. 36, Apr 2022

    work page 2022

  7. [7]

    Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of Parkinson’s disease - npj Parkinson’s Disease,

    J. Zhang, “Mining imaging and clinical data with machine learning approaches for the diagnosis and early detection of Parkinson’s disease - npj Parkinson’s Disease,”Nature, 2022

    work page 2022

  8. [8]

    Utility of multi-modal mri for differentiating of parkinson’s disease and progressive supranuclear palsy using machine learning,

    A. S. Talai, J. Sedlacik, K. Boelmans, and N. D. Forkert, “Utility of multi-modal mri for differentiating of parkinson’s disease and progressive supranuclear palsy using machine learning,”Frontiers in Neurology, vol. 12, p. 648548, 2021

    work page 2021

Show all 35 references
  1. [9]

    Machine learning classification of functional neurological disorder using structural brain mri features,

    C. Westlin, A. J. Guthrie, S. Paredes-Echeverri, J. Maggio, S. Finkel- stein, E. Godena, D. Millstein, J. MacLean, J. Ranford, J. Freeburn, C. Adams, C. Stephen, I. Diez, and D. L. Perez, “Machine learning classification of functional neurological disorder using structural bra...

    2024

  2. [10]

    Machine learning-based framework for differential diagnosis between vascular dementia and alzheimer’s disease using structural mri features,

    Y . Zheng, H. Guo, L. Zhang, J. Wu, Q. Li, and F. Lv, “Machine learning-based framework for differential diagnosis between vascular dementia and alzheimer’s disease using structural mri features,”Fron- tiers in Neurology, vol. 10, 2019

    2019

  3. [11]

    Machine-learning classi- fication using neuroimaging data in schizophrenia, autism, ultra-high risk and first-episode psychosis,

    W. Yassin, H. Nakatani, Y . Zhu, M. Kojima, K. Owada, H. Kuwabara, W. Gonoi, Y . Aoki, H. Takao, T. Natsubori, N. Iwashiro, K. Kasai, Y . Kano, O. Abe, H. Yamasue, and S. Koike, “Machine-learning classi- fication using neuroimaging data in schizophrenia, autism, ultra-high ris...

    2020

  4. [12]

    High accuracy diagnosis for mri imaging of alzheimer’s disease using xgboost,

    E. M. Arabi, A. S. Mohra, and K. S. Ahmed, “High accuracy diagnosis for mri imaging of alzheimer’s disease using xgboost,”The Open Biotechnology Journal, vol. 16, 2022

    2022

  5. [13]

    Multi-channel deep model for classification of alzheimer’s disease using transfer learning,

    S. Dharwada, J. Tembhurne, and T. Diwan, “Multi-channel deep model for classification of alzheimer’s disease using transfer learning,” in Distributed Computing and Intelligent Technology, R. Bapi, S. Kulka- rni, S. Mohalik, and S. Peri, Eds. Cham: Springer International Publis...

    2022

  6. [14]

    Multimodal attention-based deep learning for alzheimer’s disease diagnosis,

    M. Golovanevsky, C. Eickhoff, and R. Singh, “Multimodal attention-based deep learning for alzheimer’s disease diagnosis,” Journal of the American Medical Informatics Association, vol. 29, no. 12, pp. 2014–2022, 09 2022. [Online]. Available: https://doi.org/10.1093/jamia/ocac168

    2014 doi

  7. [15]

    Multimodal transformer network for incomplete image generation and diagnosis of alzheimer’s disease,

    X. Gao, F. Shi, D. Shen, and M. Liu, “Multimodal transformer network for incomplete image generation and diagnosis of alzheimer’s disease,” Computerized Medical Imaging and Graphics, vol. 110, p. 102303, 2023

    2023

  8. [16]

    A survey on mixture of experts,

    W. Cai, J. Jiang, F. Wang, J. Tang, S. Kim, and J. Huang, “A survey on mixture of experts,” 2024

    2024

  9. [17]

    Considering REM sleep behavior disorder in the management of parkinson’s disease,

    M. Figorilli, M. Meloni, G. Lanza, E. Casaglia, R. Lecca, F. L. Saibene, P. Congiu, and M. Puligheddu, “Considering REM sleep behavior disorder in the management of parkinson’s disease,”Nat. Sci. Sleep, vol. 15, pp. 333–352, May 2023

    2023

  10. [18]

    A neurologist’s guide to rem sleep behavior disorder,

    A. Roguski, D. Rayment, A. L. Whone, M. W. Jones, and M. Rolinski, “A neurologist’s guide to rem sleep behavior disorder,”Frontiers in Neurology, vol. 11, 2020

    2020

  11. [19]

    Rem sleep and neurodegeneration,

    L. P ´erez-Carbonell and A. Iranzo, “Rem sleep and neurodegeneration,” Journal of Sleep Research, vol. n/a, no. n/a, p. e14263. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1111/jsr.14263

    doi

  12. [20]

    Interhemispheric functional and structural disconnection in alzheimer’s disease: A combined resting-state fmri and dti study,

    Z. Wang, J. Wang, H. Zhang, R. Mchugh, X. Sun, K. Li, and Q. X. Yang, “Interhemispheric functional and structural disconnection in alzheimer’s disease: A combined resting-state fmri and dti study,” PLOS ONE, vol. 10, no. 5, pp. 1–16, 05 2015

    2015

  13. [21]

    T1ρand T2ρMRI in the evaluation of parkinson’s disease,

    I. Nestrasil, S. Michaeli, T. Liimatainen, C. E. Rydeen, C. M. Kotz, J. P. Nixon, T. Hanson, and P. J. Tuite, “T1ρand T2ρMRI in the evaluation of parkinson’s disease,”J. Neurol., vol. 257, no. 6, pp. 964–968, Jun. 2010

    2010

  14. [22]

    Diffusion magnetic resonance imaging-based biomarkers for neurodegenerative diseases,

    K. Kamagata, C. Andica, A. Kato, Y . Saito, W. Uchida, T. Hatano, M. Lukies, T. Ogawa, H. Takeshige-Amano, T. Akashi, A. Hagiwara, S. Fujita, and S. Aoki, “Diffusion magnetic resonance imaging-based biomarkers for neurodegenerative diseases,”Int. J. Mol. Sci., vol. 22, no. 10,...

    2021

  15. [23]

    Multi-modal brain MRI in subjects with PD and iRBD,

    S. Mangia, A. Svatkova, D. Mascali, M. J. Nissi, P. C. Burton, P. Bednarik, E. J. Auerbach, F. Giove, L. E. Eberly, M. J. Howell, I. Nestrasil, P. J. Tuite, and S. Michaeli, “Multi-modal brain MRI in subjects with PD and iRBD,”Front. Neurosci., vol. 11, Dec. 2017

    2017

  16. [24]

    Neuroimaging in the early diagnosis of neurodegener- ative disease,

    A. J. Stoessl, “Neuroimaging in the early diagnosis of neurodegener- ative disease,”Transl. Neurodegener., vol. 1, no. 1, p. 5, Jan. 2012

    2012

  17. [25]

    Early prediction of alzheimer’s disease and related dementias using real-world electronic health records,

    Q. Li, X. Yang, J. Xu, Y . Guo, X. He, H. Hu, T. Lyu, D. Marra, A. Miller, G. Smith, S. DeKosky, R. D. Boyce, K. Schliep, E. Shenkman, D. Maraganore, Y . Wu, and J. Bian, “Early prediction of alzheimer’s disease and related dementias using real-world electronic health records,...

    2023

  18. [26]

    A deep learning approach for predicting multiple sclerosis,

    E. R. Ponce de Leon-Sanchez, O. A. Dominguez-Ramirez, A. M. Herrera-Navarro, J. Rodriguez-Resendiz, C. Paredes-Orta, and J. D. Mendiola-Santiba˜nez, “A deep learning approach for predicting multiple sclerosis,”Micromachines, vol. 14, no. 4, 2023. [Online]. Available: https://w...

    2023

  19. [27]

    Multimodal deep learning for integrating chest radiographs and clinical parameters: A case for transformers,

    F. Khader, G. M ¨uller-Franzes, T. Wang, T. Han, S. Tayebi Arasteh, C. Haarburger, J. Stegmaier, K. Bressem, C. Kuhl, S. Nebelung, J. N. Kather, and D. Truhn, “Multimodal deep learning for integrating chest radiographs and clinical parameters: A case for transformers,” Radiolo...

    2023

  20. [28]

    Enhancing early parkinson’s disease detection through multimodal deep learning and explainable AI: insights from the PPMI database,

    V . Dentamaro, D. Impedovo, L. Musti, G. Pirlo, and P. Taurisano, “Enhancing early parkinson’s disease detection through multimodal deep learning and explainable AI: insights from the PPMI database,” Sci. Rep., vol. 14, no. 1, p. 20941, Sep. 2024

    2024

  21. [29]

    Deep learning integrates histopathology and proteogenomics at a pan-cancer level,

    J. M. Wang, R. Hong, E. G. Demicco, J. Tan, R. Lazcano, A. L. Moreira, Y . Li, A. Calinawan, N. Razavian, T. Schraink, M. A. Gillette, G. S. Omenn, E. An, H. Rodriguez, A. Tsirigos, K. V . Ruggles, L. Ding, A. I. Robles, D. R. Mani, K. D. Rodland, A. J. Lazar, W. Liu, D. Feny¨...

    2023

  22. [30]

    Towards understanding mixture of experts in deep learning,

    Z. Chen, Y . Deng, Y . Wu, Q. Gu, and Y . Li, “Towards understanding mixture of experts in deep learning,” 2022. [Online]. Available: https://arxiv.org/abs/2208.02813

    2022 arXiv

  23. [31]

    Med-moe: Mixture of domain-specific experts for lightweight medical vision- language models,

    S. Jiang, T. Zheng, Y . Zhang, Y . Jin, L. Yuan, and Z. Liu, “Med-moe: Mixture of domain-specific experts for lightweight medical vision- language models,” 2024

    2024

  24. [32]

    Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts,

    S. Yun, I. Choi, J. Peng, Y . Wu, J. Bao, Q. Zhang, J. Xin, Q. Long, and T. Chen, “Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  25. [33]

    Reduced volume of the putamen in rem sleep behavior disorder patients,

    T. M. Ellmore, A. J. Hood, R. J. Castriotta, E. F. Stimming, R. J. Bick, and M. C. Schiess, “Reduced volume of the putamen in rem sleep behavior disorder patients,”Parkinsonism & Related Disorders, vol. 16, no. 10, pp. 645–649, 2010

    2010

  26. [34]

    Dynamic image for 3d mri image alzheimer’s disease classification,

    X. Xing, G. Liang, H. Blanton, M. U. Rafique, C. Wang, A.-L. Lin, and N. Jacobs, “Dynamic image for 3d mri image alzheimer’s disease classification,” inComputer Vision – ECCV 2020 Workshops, A. Bartoli and A. Fusiello, Eds. Cham: Springer International Publishing, 2020, pp. 355–364

    2020

  27. [35]

    A comprehensive study of alzheimer’s disease classification using convolutional neural networks,

    Z. Guan, R. Kumar, Y . R. Fung, Y . Wu, and M. Fiterau, “A comprehensive study of alzheimer’s disease classification using convolutional neural networks,” 2019. [Online]. Available: https://arxiv.org/abs/1904.07950

    2019 arXiv

Pith tools

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