List of Papers Browse by Subject Areas Author List
Abstract
Normative modeling quantifies individual-level deviations from healthy population variability, yet its application to EEG remains limited by two fundamental gaps: existing approaches treat electrophysiological features as independent univariate variables, discarding structured inter-feature dependencies; and stochastic generative models such as Variational Autoencoders (VAEs) cannot guarantee a deterministic, invertible mapping from observations to latent space, causing the same subject to yield different deviation scores across inference runs, a critical flaw for clinical reproducibility. We propose Structure-Aware Continuous Flow Matching (SA-CFM), which addresses both gaps via physiologically grounded tokenization of EEG microstate features and a Transformer-based conditional flow matching objective. The deterministic invertibility of the learned flow yields unique, reproducible subject-level deviation representations without disease labels, a property that stochastic alternatives cannot provide by construction. Evaluated across two independent Parkinson’s disease cohorts under strictly subject-independent protocols, SA-CFM achieves unsupervised cognitive subtype separation of ΔMoCA up to 3.96 and diagnostic accuracy of 83.0% and 88.6% on two independent cohorts, substantially exceeding both end-to-end EEG classifiers and normative baselines, establishing EEG microstate normative modeling as a viable complement to neuroimaging approaches for characterizing functional network alterations in neurodegeneration. Code is available at https://github.com/gjyradl/SA-CFM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2408_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/gjyradl/SA-CFM
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuShi_SACFM_MICCAI2026,
author = { Hu, Shiang AND Hu, Yifan AND Shi, Zhuofan AND Lv, Zhao},
title = { { SA-CFM: Structure-Aware Continuous Flow Matching for Normative Modeling of EEG Microstate } },
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
This paper proposes a conditional flow matching framework for Parkinson disease diagnosis and subtype clustering. Through the deviation from reconstructed EEG microstate features in PD group and HC group, the authors quantify the atypicality of features in the patients. Introducing a KAN layer in the decoding model further improves the performance.
- 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.
In my view, the major strength of the paper is to introduce a framework of normative modeling with flow matching. Diagnosis and subtyping with the gap between reconstructed features from flow matching are interesting.
- 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.
The comparison baselines seem not to be very comprehensive. For example, what about the performance of conditional diffusion models? What about microstate features + strong classifiers (XGBoost / Transformer et al. ; The method of “Raw features” in Table 2 is not described in detail)?
- 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
I have some further questions about the method:
1.The authors used 6 datasets for training. Are the channel layouts in these datasets the same? How to deal with different channel layouts?
2.How are the microstate topographies aggregated across different subjects and datasets? Is it through hierarchical clustering? Could the authors make the microstate computation more reproducible (The code may also be made publicly available)?
3.Are the comparisons with cVAE and GAMLSS fair under supervised settings? How are they used in supervised settings?
4.The low performance of EEGNet to MedGNN is a bit surprising. Are the data properly normalized before being input to the networks?
- 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 framework is interesting and useful, but the evaluation is not comprehensive, and some important details about the method are missing. I would raise my score if the authors addressed all my concerns.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The authors have addressed all my questions.
Review #2
- Please describe the contribution of the paper
SA-CFM applies continuous flow matching to normative modeling of structured EEG microstate features. Unlike VAE-based normative models, which produce stochastic and therefore irreproducible deviation scores, flow matching is deterministically invertible — the same input always maps to the same deviation. The physiologically grounded tokenization groups microstate statistics into functional tokens before self-attention, preserving inter-microstate coupling that univariate normative approaches discard. Together, these yield reproducible, structured deviation representations that outperform both normative baselines and end-to-end EEG classifiers on unsupervised cognitive subtyping and diagnostic accuracy in PD. The problem framing and technical solution are both well-motivated.
- 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 case against stochastic normative models is principled and correct. VAE posterior sampling introduces irreducible variance into deviation scores, and using a deterministic flow is a well-motivated solution. The evaluation is methodologically rigorous: two independent PD cohorts, strictly subject-independent 5-fold MCCV with fixed seeds — rare in EEG-based neuroimaging. The ablation is clean; Table 4 shows each component contributing incrementally, with the full model consistently outperforming all ablations across both cohorts. The interpretability analysis holds up: the dissociation between transition-specific discriminability (Trans_A_D, Trans_A_C) and global entropic signatures is a biologically meaningful finding, not post-hoc decoration. Training exclusively on 1,419 healthy controls with no disease labels is the correct normative paradigm. And the near-chance performance of end-to-end EEG classifiers is reported honestly and used to justify the normative framing, which strengthens rather than undermines the argument.
- 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- AEPD has only N=32 patients! the second independent cohort on which the generalization claim partly rests. Under 5-fold MCCV, some folds have fewer than 7 test subjects. The standard deviations are large (±6.48%), and the reported 88.6% subject-level accuracy on this cohort is unreliable. Conclusions drawn from AEPD should be heavily caveated. 2- The normative training data is heterogeneous and unharmonized. The 1,419 healthy controls span age 17–88 across five datasets with different acquisition protocols. EEG microstate statistics are acquisition-sensitive, and pooling without site correction could introduce systematic biases into the normative reference. 3- The KAN-VAE downstream head is unjustified. A VAE is used as the classification head on top of the flow matching normative model, reintroducing stochastic sampling at the exact stage the paper argues against. The justification ( that a VAE induces a structured latent space for clustering ) is weak and directly contradicts the central argument for deterministic reproducibility. 4- Only age is conditioned on. Sex, education, medication status, and disease duration (all known confounds in PD cognitive research ) are ignored. This is a significant limitation that goes undiscussed. 5- No reproducibility analysis is reported. The central empirical claim is deterministic reproducibility over VAE stochastic inference — yet no quantitative comparison of run-to-run deviation score variance between SA-CFM and cVAE is provided. This is the most important claim in the paper and it is supported only theoretically.
- 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.
- 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?
It is a well-motivated paper with a clear contribution. The deterministic invertibility argument is principled, the ablation is convincing, and dual-cohort validation is commendable. The weaknesses are real — the AEPD cohort is too small to support strong generalization claims, the reproducibility analysis is missing, the KAN-VAE head is inconsistent with the central argument, and single-covariate conditioning is underexamined — but none are fatal.
- Reviewer confidence
Somewhat confident (2)
- [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.
The rebuttal resolves two of my five concerns. The KAN-VAE inconsistency is addressed through clarification; the determinism claim applies to the flow inversion stage, not the downstream head, and the controlled comparison is fair. The single-covariate conditioning is a genuine data constraint, not a design choice. The AEPD cohort size (N=32) is appropriately conceded and reframed as secondary evidence. The heterogeneous normative data concern is reasonably mitigated by the unified preprocessing pipeline. The absence of a quantitative reproducibility analysis remains a weakness of the submitted paper; the central empirical claim that deterministic inversion outperforms stochastic VAE inference is supported theoretically but not empirically demonstrated. This was a valid concern at submission and remains so. I maintain my score of 4.
Review #3
- Please describe the contribution of the paper
1.A flow matching framework with physiologically grounded tokenization that models microstate features as interacting structural units rather than flat vectors. 2.An empirical demonstration that deterministic flow inversion yields more stable and reproducible deviation representations than stochastic VAE-based normative modeling. 3.Validation on two independent PD cohorts, achieving superior unsupervised cognitive subtyping and diagnostic classification compared to both normative baselines and end-to-end classifiers.
- 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.
1) The paper is very well written 2) The exploration of microstates constitutes both the primary novelty and the core strength of this paper. 3) The framework is feature-agnostic and inherently stable.
- 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.
None
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
None
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This research introduces a highly innovative framework that represents a substantial leap forward for the discipline.
By moving beyond conventional methodologies, the authors have successfully addressed long-standing technical limitations.
The proposed approach establishes a new benchmark, offering unprecedented analytical capabilities and structural efficiency.
Furthermore, this advancement equips both researchers and industry practitioners with a more robust foundation for practical application.
Ultimately, the paper’s novel contributions establish a compelling new trajectory for subsequent research and development in the field.
- 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.
This research introduces a highly innovative framework that represents a substantial leap forward for the discipline.
By moving beyond conventional methodologies, the authors have successfully addressed long-standing technical limitations.
The proposed approach establishes a new benchmark, offering unprecedented analytical capabilities and structural efficiency.
Furthermore, this advancement equips both researchers and industry practitioners with a more robust foundation for practical application.
Ultimately, the paper’s novel contributions establish a compelling new trajectory for subsequent research and development in the field.
Author Feedback
We thank the reviewers and AC for the constructive comments. We focus on the key concerns raised in the meta-review and reviews. [AC/Meta] We address the main concerns identified in the meta-review: reproducibility, baseline fairness, evaluation completeness, cohort size, data heterogeneity, and methodological details. [R1] Baselines and fairness. cVAE is the direct stochastic conditional-generative baseline for testing deterministic normative inference; a conditional DDPM would also be stochastic and would not directly test deterministic inversion. Microstate+XGBoost/Transformer classifiers test supervised discriminative learning rather than normative deviation modeling. Our controlled comparison fixes the 27-D input and the same KAN-VAE head for GAMLSS, cVAE, and SA-CFM, isolating deviation encoding. “Raw Features” in Table 2 means k-means directly on the 27-D vector, a non-normative lower bound. [R1] Channel layout and microstate computation. All datasets were standardized to the common 19-channel 10-20 montage; for recordings with more channels, only these shared 19 channels were retained. The model consumes 27-D microstate statistics, not channel-indexed signals. GFP-peak k-means (k=4, 10 restarts) is followed by spatial-correlation matching to canonical A-D templates on this shared montage; statistics are computed in this fixed order. No hierarchical clustering is used. [R1] Low end-to-end EEG accuracy. Inputs were z-scored per channel per subject. Under subject-independent splits, the near-chance results of five end-to-end baselines in Table 3 suggest that direct segment-level discrimination is unstable under strong inter-subject heterogeneity, motivating normative deviation modeling. [R2] Reproducibility of deviation scores. We distinguish inference determinism from training randomness. For a fixed trained SA-CFM model and fixed-step RK4 solver, backward ODE integration maps each input to one unique latent deviation representation; no posterior sampling is involved. In contrast, cVAE inference uses a posterior distribution and, even with the 10-MC averaging in Sec. 3.2, retains sampling-induced variance. Table 2 reports clustering stability under the same downstream head: SA-CFM ARI = 0.58/0.65 vs cVAE 0.33/0.39.[R2] AEPD size and generalization. We agree that AEPD (N=32) should be read as a secondary independent-cohort test, not the sole basis for generalization. The primary evidence is PD-RS (N=100) under subject-independent 5-fold MCCV with 5 fixed seeds, and AEPD uncertainty is reported transparently. The claim rests on cross-cohort consistency: Subj-Acc 83.0%/88.6%, deltaMoCA 3.32/3.96, and ARI 0.58/0.65.[R2] Heterogeneous normative data. All training cohorts used the same pipeline (128 Hz, 0.5-45 Hz filtering, ADJUST-ICA, GFP-peak k-means). We use 27-D microstate temporal/transition statistics rather than channel-indexed raw signals, and condition on age, the dominant covariate available in all cohorts. Fig. 2(a) shows a smooth unimodal latent norm distribution with PD-RS HCs overlapping the normative range, suggesting no gross site-separated structure. [R2] KAN-VAE and determinism. The deterministic claim concerns the normative deviation representation before any downstream head. KAN-VAE is applied identically to GAMLSS, cVAE, and SA-CFM, so its stochasticity is controlled and does not alter SA-CFM inversion in Eq. 5.Table 4 shows SA-CFM without KAN already improves over naive/attention CFM; KAN is an additional shared nonlinear readout. [R2] Single-covariate conditioning. Age is the only covariate uniformly available across all six datasets. Sex, education, and medication are missing in several training cohorts; including them would shrink or bias the normative reference. Disease duration cannot be used for a model trained exclusively on healthy controls. We hope these clarifications address the reviewers’ and AC’s major concerns.
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 summarized strengths of your work, novel aspects and noted that your solution may be of potential interest. They all agree on some novelty of this research, namely that they find this a well-motivated paper with clear contribution, convincing ablation study, and interesting results of diagnosis and subtyping. They also find that the paper is in principle well written. However, reviewers also point in detail to some major weaknesses of the current work as they have questions on baseline comparisons, that the evaluation does not seem comprehensive, and that important details on the method are missing. There are questions on the small size of the EPD cohort towards generalization, and reviewers also find that details for reproducibility analysis are missing. In summary, we see that weaknesses and limitations somewhat balance its merits and thus may pose questions for this work to be accepted for MICCAI. Would this paper be invited for rebuttal, please focus your answer on the major limitations and weaknesses found by reviewers which would be most decisive for decision making in a concise, clear way.
- 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 think that the idea is interesting. Although some weaknesses still remain, I thought that authors have provided a good discussion on this early-stage but interesting research, and have provided some convincing responses to specific critiques. I strongly encourage authors to include better clarifications and revisions as suggested by reviewers to strengthen the final version. I would argue for accept.
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 study presents a technically innovative method for normative modeling of EEG, validated on cohorts with acceptable sample sizes. The novelty outweighs the limitation in experiments.
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.
Even though one of the reviewers still has some concerns, most of the initial issues have been addressed in the rebuttal, and all reviewers now recommend acceptance. I agree with this recommendation.