List of Papers Browse by Subject Areas Author List
Abstract
Neurodegenerative diseases exhibit substantial heterogeneity, complicating both diagnosis and prognosis. Identifying clinically meaningful subtypes is crucial for understanding disease mechanisms and improving diagnostic precision and prognostic accuracy. Existing subtyping approaches primarily rely on unsupervised learning of patient data for capturing inter-individual variability, but they may fail to uncover subtypes that are directly informative for diagnosis or prognosis. To address this limitation, we propose a novel mixture-of-experts (MoE) framework that integrates predictive modeling with subtype discovery. Unlike traditional subtyping methods, our approach learns a router to assign individuals to specialized expert networks, each corresponding to a distinct subtype, for improving predictive accuracy. By coupling subtype identification with prediction, the proposed MoE framework encourages the discovery of subtypes that are not only statistically distinct but also clinically informative. We evaluate the framework on a real-world dataset of mild cognitive impairment (MCI) subjects and a semi-simulated dataset, demonstrating superior performance in predicting progression from MCI to Alzheimer’s disease while revealing distinct, clinically meaningful MCI subtypes. Code is available at https://github.com/Kateridge/MoESubtyping.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0454_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Kateridge/MoESubtyping
Link to the Dataset(s)
ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/
OASIS-3 dataset: https://sites.wustl.edu/oasisbrains/home/oasis-3/
AIBL: https://aibl.org.au/
BibTex
@InProceedings{ZhaYua_Discovering_MICCAI2026,
author = { Zhang, Yuanwang AND Li, Hongming AND Fan, Yong},
title = { { Discovering Heterogeneous Neurodegenerative Disease Patterns From MRI Data for Improved Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
By introducing a mixture-of-experts framework, the paper detects subtypes by explicitly distinguishing them during learning. They apply their method to a real-world dataset to predict the progression to Alzheimer’s disease in patients who are initially at risk.
- 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.
The major strength of the paper is that the proposed method is simple and generalizable: it proposes to add a mixture of expert models after an MLP acting as a router to different experts. The method beats what is generally considered state-of-the-art for tabular data, namely, random forests. The paper is well written, and the ablation study is well performed.
- 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.
One “weakness” is that the method is applied only on tabular data: the authors should discuss how we can apply such a framework to a representation space coming from self-supervised learning of imaging data. The other weakness is the relationship between the experts and the actual labels: the loss (1) forces the subjects having the same subtype label to belong to the same cluster. However, there is a lengthy discussion on the optimal number of experts: is it not expected that it equals the number of subtypes (3 if I am not mistaken)? When the authors state that there are 4 experts (figure 3 of page 8), the authors should add the proporition of the actual subtypes for each expert for clarity.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
P. 5 How is done the semi-simulated dataset? The authors should be more precise that just saying “by manually introducing distinct atrophy patterns” P. 6 By comparing with random forests, can you write how is done the hyperparameter search for Random Forest? p. 6: I suppose that the tabular foundation model has been used as zero-shot: can you add this piece of information for clarity? Table 2: no standard deviation is given, which makes comparison between models hard
- 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?
The method is simple and generalizable and is of general audience. The method is sound and is presented clearly. I hope that it may trigger the authors or others to apply this method combined with self-supervised learning on imaging data
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I suggested an accept in my initial review. I stick to this. I believe the method is of general interest even if it is applied only on tabular data.
Review #2
- Please describe the contribution of the paper
The paper proposes a novel mixture-of-experts (MoE) framework to jointly perform prediction and subtype identification for neurodegenerative diseases using MRI data.
- 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 integration of prediction and clustering via a MoE architecture tailored for heterogeneous neurodegenerative disease data.
-
Use of multiple regularization losses to stabilize training and prevent degenerate solutions.
-
Comparisons to strong baselines and state-of-the-art models in classification and survival prediction tasks.
-
- 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.
-
Dataset and MCI case selection: The paper states that for classification, 668 MCI subjects with baseline MRI were selected and categorized as stable or progressive based on dementia progression within three years. For progression risk prediction, 1,243 MCI subjects with baseline MRI and multiple follow-ups were used. It is not clear how these subjects were selected.
-
Optimal cluster number determination: The authors report that performance peaks at four experts, suggesting four natural subtypes. However, relying exclusively on prediction metrics to select cluster number is weak (the difference in performance is extremely narrow) which may not fully capture the biological or clinical validity of subtypes
-
Clustering evaluation: Clustering performance is quantitatively assessed on the semi-simulated dataset using ARI and NMI, showing improvements with the MoE framework. However, the real-world clustering quality is not directly evaluated, and no comparison to alternative subtyping methods is provided in the real data setting.
-
Comparison to baselines: The paper compares predictive performance against baselines for classification/risk prediction. However, comparisons against other subtyping or clustering strategies, especially those designed for neurodegenerative disease heterogeneity, are missing. This would clarify the added value of the MoE approach for subtype discovery.
-
Interpretation of atrophy patterns: The four identified subtypes show distinct atrophy patterns consistent with disease stages (e.g., mild atrophy in early MCI subtype, severe atrophy in late-stage subtype). While this supports clinical plausibility, it raises the question of whether the method primarily captures disease severity rather than novel subtype heterogeneity. Further exploration of subtype-specific clinical or biomarker profiles would enhance interpretability.
-
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed approach is potentially interesting, but improvement over existing clustering/subtyping methods is unclear.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
There are some methodological and conceptual inconsistencies and it’s not clear what the work is adding to existing methods.
Review #3
- Please describe the contribution of the paper
Neurodegenerative diseases exhibit substantial heterogeneity, making it difficult to identify subtypes that are both biologically meaningful and clinically useful for diagnosis and prognosis. To address this, the manuscript proposes a mixture-of-experts framework that jointly performs predictive modeling and subtype discovery by learning to assign individuals to specialized expert networks representing distinct subtypes. Evaluated on real-world and semi-simulated MCI datasets, the approach improves prediction of progression to Alzheimer’s disease while identifying clinically meaningful subtypes.
- 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.
- very important/ timely topic
- manuscript is easy to understand
- experiments are well thought out -Display of subtypes in Figure 3
- 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.
- fails to account for confounders such as sex or age so subtype might be simply capturing that
- Is difference of more MOE statistical significant over single ?
- Does not seem much better than TabFPN or Random Forest
- Please report balanced accuracy
- Not clear from description how method enforces to have distinct experts
- why does accuracy peak at 4 experts ? What is captured in 4 experts that is not in 2 or 3 ?
- Application is based on tabular data vs medical images
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
interesting submission but lacks depth and rigor to be of higher impact
- Reviewer confidence
Very confident (4)
- [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 reviewers for their constructive feedback. Following meta-review, we focus on main concerns and then address reviewer-specific comments.
Q1: Expert/subtype interpretation, severity, and confounders [R2,R3,Meta] We agree that learned experts should not be overinterpreted as biological subtypes without further validation. We interpret them as data-driven disease groups. With a small K (e.g., 2), the model may capture a mild–severe atrophy axis, leading to poorer prediction performance (Table 1). With K=4, experts show spatially distinct atrophy patterns, including medial-temporal-predominant and diffuse cortical patterns, suggesting heterogeneity beyond severity alone (Fig. 3). Assessing clinical records of Subtype 1 (S1) and S3 showed similar conversion rates (~20%) and comparable cognitive decline, despite their distinct spatial patterns. A similar pattern was observed for S2 and S4, which both showed ~50% conversion and comparable cognitive decline. Including age and sex as additional inputs yielded similar prediction performance. Nevertheless, we agree that explicit age/sex comparisons across subtypes and additional clinical profiling would further strengthen interpretation and is an important future work.
Q2: Experts number K [R1-3,Meta] Since no ground-truth disease subtype labels exist, we use prediction performance as a surrogate indicator. We agree it is insufficient alone for selecting K when differences are small. Evaluating current subtyping results with clustering stability (ARI) and quality (Calinski-Harabasz Index) provided additional insight (K=2-5; ARI: 0.896, 0.840, 0.822, 0.525; CH: 195.2, 113.4, 116.1, 98.8, respectively). Although K=2 was most stable, it mainly separated mild from severe atrophy. In contrast, K=4 provided the best balance among prediction performance, stability, and heterogeneity.
Q3: Expert separation and routing enforced [R1,R3,Meta] Each expert represents one subtype, while the router learns subtype assignments that minimize prediction loss. Expert separation is encouraged by a load-balancing term preventing collapse to one expert and sparse routing to assign each subject to a single expert. Thus, experts are encouraged to specialize while maintaining balanced usage.
Q4: Evaluation relative to baselines and clustering methods [R2,R3,Meta] Our goal is not standalone clustering, but joint learning of subtypes and prediction in a unified framework that is not hinged on specific prediction or clustering backbones. Table 1 shows that, even with simple MLP experts, our method is competitive with or improves over TabPFN and DeepSurv. Table 2 shows that MoE improves clustering performance over alternatives. In response to R3, MoE significantly improves over Single and Single*, with all p-values <0.01.We agree that additional comparison with other methods on real data would further strengthen biological validation, but real-world subtype labels are unavailable; therefore, following prior studies [7,24,25,26], we used semi-simulated data with known subtype structure for quantitative clustering evaluation.
Q5: Additional Clarification [R2] Risk prediction used 1,243 MCI subjects with baseline MRI and longitudinal follow-up. For classification, stable/progressive MCI was defined by conversion to dementia within 3 years; subjects without sufficient follow-up were excluded, yielding 668 subjects. [R1] Semi-simulated atrophy patterns were generated by reducing selected ROI values by predefined percentages to mimic subtype-specific neurodegeneration. [R1,R3] We used ROI-level MRI features for interpretability and controlled semi-simulation, but the framework is representation-agnostic and can operate on voxel-wise imaging embeddings (discussed in P.8). [R1,R3] Random Forest hyperparameters were tuned using sklearn RandomizedSearchCV; TabPFN was evaluated zero-shot; standard deviations for Table 2 and balanced accuracy can be added. We will revise the paper accordingly for better clarity.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Reviewers mainly request clarifications on whether the identified experts truly reflect meaningful biological subtypes rather than confounders or disease severity, and why the model consistently favors four experts. They also ask for clearer explanations of the method (how expert separation is enforced and how routing works), along with clarification of evaluation choices, such as metrics, and how improvements should be interpreted relative to existing baselines.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
I tend to agree with Reviewer 2’s assessment. The article presents a relatively straightforward idea, combining subgroup discovery with supervised classification, that has already been explored in the literature. In its current form, the manuscript does not provide a sufficiently thorough comparison with prior work, either from the perspective of subgroup identification or predictive modeling.
In particular, the proposed approach should be compared against established subgroup discovery methods such as HYDRA, CHIMERA, UCSL, Deep UCSL, MAGIC, SuStaIn, and DeepCluster (or related deep clustering approaches). Such comparisons are essential to demonstrate whether the proposed method offers any advantage in terms of subgroup stability, interpretability, clinical relevance, or predictive performance.
Similarly, stronger baselines should be included for the supervised classification and progression-risk prediction tasks. Without these comparisons, it is difficult to assess the true contribution of the method.
More fundamentally, I share Reviewer 2’s concern that the methodological novelty of the work remains unclear. The manuscript does not convincingly explain what distinguishes the proposed framework from existing approaches, nor why it should be expected to outperform them either theoretically or empirically.
Finally, the evaluation would benefit from the inclusion of an additional dataset, ideally one in which subgroup labels or underlying subgroup structure are known (or can be reasonably approximated). This would provide stronger evidence that the method is capable of recovering meaningful subgroups rather than dataset-specific patterns or task difficulty.
Overall, the clinical utility and statistical added value of the proposed approach remain insufficiently demonstrated. The manuscript would benefit from a clearer positioning with respect to the existing literature, stronger baselines, and more rigorous validation of the discovered subgroups.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have reasonably clarified the main concerns about expert interpretation, expert number selection, and routing mechanism; while subtype validation remains limited, the revised explanation is sufficient to support acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I recommend accepting this paper. The idea to couple identification of data-driven disease groups with predictive modelling in a Mixture-of-Experts framework is novel and interesting. Although the paper does not provide sufficient validation to claim that each expert discovers a distinct disease subtype rather then disease severity stage, during rebuttal authors provide disease progression comparison for identified groups, which do not show a linear trend. The identification of data-driven disease groups and their processing through separate experts improve predictive performance over a single-expert approach as well as other baseline methods, demonstrating that the proposed approach is both novel and effective.