Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Preclinical Alzheimer’s Disease (AD) detection remains challenging due to the complex interplay of biological, structural, and temporal factors. Existing methods often struggle to integrate multimodal longitudinal data and predict key clinical outcomes. We propose MAGNET-AD, a novel multitask spatiotemporal graph neural network designed to predict the Preclinical Alzheimer’s Cognitive Composite (PACC) score and time to AD conversion. MAGNET-AD offers three key contributions: (1) A dynamic heterogeneous graph architecture with weighted edges for hybrid fusion mechanisms, integrating static and dynamic multimodal data; (2) a temporal importance weighting loss function that adaptively learns critical time points while jointly optimizing time prediction and cognitive decline estimation; and (3) an interpretable attention framework that highlights key brain regions and genetic factors driving disease progression. MAGNET-AD achieves state-of-the-art performance with a concordance index of 0.858 for conversion time prediction and a mean square error of 1.983 for PACC prediction, outperforming existing deep learning approaches. These results underscore MAGNET-AD’s potential for early AD risk assessment and monitoring, enabling broader clinical applications. The code and processed data will be available at https://github.com/BioMedIA-MBZUAI/MAGNET-AD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0447_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/BioMedIA-MBZUAI/MAGNET-AD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{HasSal_MAGNETAD_MICCAI2025,
        author = { Hassan, Salma AND Salem, Mostafa AND Papineni, Vijay Ram Kumar AND Elsayed, Ayman AND Yaqub, Mohammad},
        title = { { MAGNET-AD: Multitask Spatiotemporal GNN for Interpretable Prediction of PACC and Conversion Time in Preclinical Alzheimer } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15973},
        month = {September},
        page = {348 -- 357}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    Authors presented a multitask spatiotemporal graph neural network to predict the PACC score and time to AD conversion. The proposed method integrates static and dynamic multimodal data using a dynamic heterogeneous graph architecture with weighted edges. They also devised a weighting loss function and an interpretable attention technique

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Leveraging spatiotemporal multimodal data for joint prediction of time-to-AD conversion and PACC score trajectories.
    • A comprehensive ablation study validating the contribution of each model component.
    • Interpretable identification of critical biomarkers, including vulnerable brain regions and genetic factors.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Evaluations were limited to conventional methods without comparison to state-of-the-art approaches.
    • Critical limitations and potential solutions to address them were not discussed.
    • Missing some details in the preprocessing steps.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    1- Some necessary details in the preprocessing steps are unclear. For example, how were the 100 AD-related genes selected, and how were the SNPs identified? 2- The method for defining hyperparameter values and their impact on model performance is not clearly explained. Please provide justification for the chosen values. 3- Comparisons with state-of-the-art methods would strengthen the evaluation, rather than only baseline approaches. Adding such comparisons would better demonstrate the paper’s contributions. 4- Equations should be properly positioned in the text (near their first mention) and accompanied by explanations of their variables. Also, use “Eq. (X)” instead of “Equation (X)” for conciseness. 5- Some abbreviations (e.g., RNN, MCI) are defined but never used, while others (e.g., BOLD) are used without definition.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The authors propose a novel multitask spatiotemporal graph neural network (MAGNET-AD) to predict the preclinical Alzheimer’s cognitive composite (PACC) score and time to AD conversion. A dynamic heterogeneous graph architecture with weighted edges for hybrid fusion mechanisms, integrating static and dynamic multimodal data. Authors validate the method on conversion time prediction and PACC prediction, outperforming baseline methods from literature.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    MAGNET-AD presents a novel multitask spatiotemporal graph neural network framework for predicting preclinical Alzheimer’s disease (AD) progression. By leveraging an innovative hybrid fusion architecture, adaptive temporal weighting, and interpretable attention mechanisms, MAGNET-AD demonstrates strong capabilities in modeling the complex interactions between genetic factors and longitudinal brain changes.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    1. The authors use SynthSeg to extract masks for 32 brain regions. It remains unclear why alternative tools that provide more fine-grained anatomical parcellation were not employed, which could potentially uncover more subtle regional associations.
    2. The construction of temporal graphs is based on radiomic features from different brain regions. However, it is unclear why the embeddings derived from the AnatCL foundational model were not directly utilized, as these may better capture anatomical consistency across time.
    3. The method illustration figure lacks clarity. It would benefit from a more structured and visually intuitive layout to help readers better understand the framework.
    4. The manuscript contains several formatting issues throughout the text, which should be carefully checked and revised for readability and professionalism.
    5. The paper refers to the first loss function, but its explicit formulation is missing. Additionally, in Equation (2), some variables are introduced without proper definition, which hinders understanding and reproducibility.
    6. It is unclear whether the baseline methods reported in Table 1 and Table 3 are identical. If so, the inconsistency in their reported performance across tables requires clarification, especially in terms of experimental setup and evaluation metrics.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The manuscript presents an innovative multitask spatiotemporal GNN framework, MAGNET-AD, which integrates dynamic and static multimodal data through a hybrid fusion mechanism and interpretable attention layers. It is evaluated on two clinically relevant tasks—PACC score prediction and AD conversion time estimation—demonstrating improved performance over baseline models. Despite these promising contributions, several issues limit the manuscript’s clarity and completeness. First, the authors use SynthSeg to generate masks for 32 brain regions but do not justify why more fine-grained anatomical atlases were not explored, which might reveal subtler regional associations. Second, it is unclear why radiomics features were chosen over directly using feature embeddings from AnatCL, which may offer richer representations. Third, the method illustration is hard to follow, and several key variables and formulas (e.g., in Equation 2) lack definition, making reproducibility difficult. Furthermore, inconsistencies between Tables 1 and 3 raise concerns about experimental clarity, and overall formatting issues affect the manuscript’s readability. Nonetheless, the proposed model introduces a novel architecture with promising results, and the integration of spatiotemporal attention and hybrid fusion presents a meaningful advancement. Therefore, I recommend Weak Accept, conditional on the authors addressing clarity, reproducibility, and presentation issues in a revision.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    MAGNET-AD is a multitask spatiotemporal graph neural network designed to predict the PACC score and time to AD conversion using multimodal longitudinal data. It integrates MRI, fMRI, genetic, and EHR data into a graph structure with static and dynamic features. Innovations include hybrid fusion, adaptive loss, and dual attention for interpretability. Evaluated on the A4 cohort, the model outperforms existing methods in accuracy and interpretability.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    Innovative architecture capturing both temporal progression and static biological information. Joint prediction of cognitive scores and disease progression increases clinical relevance. Using a cognitive composite is more sensitive and clinically relevant than relying on single tests. Dual attention mechanisms provide meaningful insights into disease-relevant brain regions and genetic markers, aligning with known pathological staging. Advancements through this work are well outlined. Strong empirical results across multiple visit and data configurations. Use of the A4 preclinical cohort and inclusion of survival analysis metrics like concordance index supports translational relevance.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Evaluated only on A4 data; external validation (e.g., ADNI, OASIS) is needed. Likewise, validation with a disease-specific metric like amyloid imaging would be useful. As presented it seems the approach could penalize different disease progression rates, which limits clinical relevance. Lacks detailed reporting of demographics, variable definitions, and EHR contributions. Complex architecture may be difficult to implement clinically. Attention-based interpretability requires further biological validation and translation. The multi-head attention, large multimodal graphs, and dual loss optimization may require substantial computational resources.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Using cognitive trajectories and temporal data, this work presents a solid contribution to early intervention modeling.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank the reviewers for their thoughtful feedback and for highlighting key strengths, including the novel multitask GNN architecture, dynamic multimodal integration, and clinically relevant joint modeling of PACC and AD conversion. We appreciate the recognition of our interpretability approach and thorough ablation analysis. We address all comments below and will revise the manuscript accordingly.

R1, R3 – Baseline Comparisons and Experimental Rigor As this is the first study to jointly model PACC trajectories and AD conversion and A4 dataset is new, there are no directly comparable benchmarks. We adapted baselines from related literature since the task and cohort are novel.

R3 Regarding the concern about inconsistencies between Table 1 and Table 3: these two tables are based on different experimental conditions. Table 1 evaluates the model using the full longitudinal data available (up to 11 visits per patient). In contrast, Table 3 simulates realistic clinical scenarios with limited follow-up (2–5 visits). The observed drop in performance is expected due to reduced temporal context. Importantly, even in these constrained settings, our model consistently outperforms all baselines.

R2 – External Validation and Dataset Generalizability We appreciate R2’s suggestion to validate using external datasets. However, our work focuses on preclinical Alzheimer’s Disease. The A4 Study is the only large-scale public dataset designed for this cohort. In contrast, ADNI, OASIS, and others focus on healthy, MCI, or AD, making them unsuitable for early intervention modeling.

R1, R3 – Preprocessing and Feature Choice Justification The 100 AD-related genes were selected from the Human Protein Atlas based on brain expression and AD relevance. SNPs were included in the A4 dataset and filtered using the chromosome and position range of the selected genes.

SynthSeg was used for segmentation due to its robustness to scanner variability and anatomical changes in elderly brains. Unlike atlas-based methods, SynthSeg does not rely on rigid templates, which often degrade in the presence of atrophy.

R3 - Both AnatCL and radiomics features were used in a complementary manner. AnatCL embeddings from temporal scans were used as node features in the spatial layer to capture anatomical identity. Using them again for temporal edges would be redundant. Instead, radiomic features, which better reflect subtle longitudinal changes, were used to define temporal edge weights.

R1 – Hyperparameter Selection and Limitations Hyperparameter tuning was conducted via a structured search. All configurations and search ranges will be included in the code release. We thank R1 for highlighting the need to discuss limitations. In the final manuscript, we will note the limited PET availability, and the opportunity to include more accessible biomarkers, such as blood-based, to improve clinical applicability.

R1, R2 – Interpretability and Clinical Relevance Our dual-attention mechanism identifies relevant genes and brain regions, aligning with known AD pathology. Further biological validation is planned. R2’s concern on penalizing varying progression rates is addressed via a DeepHit-based survival component, which models individualized risk distributions. We also recognize that the model’s complexity may pose clinical integration challenges. Future work will explore simplified and modular deployment options tailored to available data and clinical priorities.

R1, R2 – Terminology and Reporting Details We will define all abbreviations at first use, remove unused terms, and clarify terminology. A4 demographics are publicly available. Variable and EHR feature descriptions will be provided in the code release. Figure layout, equation positioning, and formatting will also be improved.

R1, R2, R3 – Reproducibility and Open Access We will release full code and preprocessing scripts upon acceptance to support reproducibility and future benchmarking.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A



back to top