List of Papers Browse by Subject Areas Author List
Abstract
Electroencephalography (EEG) foundation models are designed to learn neural representations that generalize across tasks, subjects, and recording conditions, reducing reliance on task-specific designs and improving scalability in real-world applications. However, existing approaches face two major challenges: limited semantic modeling, which restricts generalization to unseen datasets, and overreliance on masked reconstruction objectives that emphasize local details while neglecting global structure. To address these issues, we propose LEGEND, a Language-aligned EEG foundation model with flow matching latENt Denoising. Our framework integrates two complementary modules: Signal–Language Alignment (SLA), which reformulates EEG decoding as instruction-based generation under natural language prompts, and Flow Matching Latent Denoising (FMLD), which promotes structural abstraction through latent-space denoising. We pretrain LEGEND on 8 public EEG datasets and evaluate it on 6 datasets for full fine-tuning, as well as on 2 unseen datasets for zero-shot classification. Results show that LEGEND achieves 15 top-1 full fine-tuning results out of 18 evaluations across 7 baselines and 3 metrics while maintaining strong zero-shot performance, demonstrating its effectiveness as a general-purpose EEG foundation model. The code is available at https://github.com/5GYYYYY/LEGEND.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3568_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/5GYYYYY/LEGEND
Link to the Dataset(s)
N/A
BibTex
@InProceedings{GuiYiy_LEGEND_MICCAI2026,
author = { Gui, Yiyu AND Chen, Mingzhi AND Luo, Guibo},
title = { { LEGEND: A Language-Aligned EEG Foundation Model with Flow Matching Latent Denoising } },
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
The paper proposes LEGEND, a novel EEG foundation model that integrates language alignment and flow matching latent denoising to improve generalization across tasks, subjects, and recording conditions.
- 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 paper proposes a novel EEG foundation model (LEGEND) that integrates language alignment with flow matching latent denoising, offering a unified framework for learning generalizable EEG representations across diverse settings.
- 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 descriptions of the two proposed models are unclear. More details are needed, and it should be clarified whether the two identified challenges are effectively addressed.
- 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 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?
The score is mainly based on the novelty and potential impact of the proposed method, balanced against the lack of clarity in model descriptions and insufficient evidence demonstrating that the identified challenges are effectively addressed.
- 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
This paper proposes a foundation model for EEG with two key innovations: alignment between EEG signals and semantic concepts, and denoising with flow matching. This goal of this is to prevent focus on spurious local signal changes and and to provide a more robust noise-invariant reconstruction, both of which should lead to better performance on new tasks in the zero and few shot settings. The paper compares the proposed method to various baselines with full fine-tuning, outperforming other methods in most cases, and tested the proposed method in a zero and one-shot setting where it also outperformed other methods.
- 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 methodological additions are particularly well-thought out. The model is also well-validated with validation on six different datasets, encompassing 8k patients, and covering different neurological diseases. Additionally, the method is compared against several baselines, which strengthens the findings of superior performance across 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.
While the paper provides several comparator baselines, there are a few important FMs in the EEG space that it does not compare against, making it unclear if LEGEND would outperform them.
- 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.
(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?
The paper presents well-thought out methodological innovations that fit the target domain and provide strong validation against various benchmarks and comparator models.
- Reviewer confidence
Somewhat confident (2)
- [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
The paper proposes a new foundation model for Electroencephalography (EEG) data. It comprises two main modules (1) the Signal–Language Alignment (SLA) (2) a Flow Matching Latent Denoising (FMLD).
- 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 main strength is the evaluation part. The paper presents subject-wise 5-fold cross-validation results on 6 datasets from different tasks and compares the proposed approach with 7 competing methods. It also includes an ablation study to assess the contributions of the two main modules (SLA and FMLD).
- 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.
I don’t think either of the two modules is new; probably, there is some novelty in the fact that they haven’t been used in this context before. It is not clear which of the competing methods are also foundational models and if any of them also combines the EEG signal with text.
- 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
Please answer to the weakness comments. Minor corrections: all the acronyms in the references need to be in upper case (eeg, bci, etc..) and a few references (9,10, etc..) are missing the date
- 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?
The authors need to be clearer on the novelty of their method. Are there any other foundation models that also include text, and how? They also need to clarify which methods in the comparison are also foundation models and on which datasets they were trained.
- Reviewer confidence
Somewhat confident (2)
- [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 sincerely thank the reviewers and the meta-reviewer for their constructive comments and positive assessment of our work.
Reviewer 1: The descriptions of the two proposed modules are unclear, and it should be clarified whether the two identified challenges are effectively addressed. A: We thank the reviewer for this helpful comment. LEGEND addresses the two identified challenges through two complementary modules. Signal-Language Alignment (SLA) tackles insufficient semantic modeling by reformulating EEG decoding as instruction-based generation: given an EEG signal and a natural-language prompt, the model generates answer tokens autoregressively, thereby aligning EEG representations with task-level semantic concepts and enabling zero-shot inference on unseen datasets. Flow Matching Latent Denoising (FMLD) addresses the overreliance on local masked reconstruction by performing denoising in the latent space rather than reconstructing raw EEG amplitudes, encouraging smoother and more invariant structural representations. The ablation study supports their complementary roles, as removing either SLA or FMLD generally degrades performance. We will clarify this module-to-challenge correspondence in the camera-ready version.
Reviewer 2: The paper may miss several important EEG foundation models, making it unclear whether LEGEND would outperform them. A: We appreciate this comment. Our comparison was intended to cover representative baselines from two categories: task-specific/end-to-end EEG decoding models, including EEGNet, EEGConformer, and SPaRCNet, and recent pretraining-based EEG foundation models, including BIOT, LaBraM, CBraMod, and CSBrain. These baselines cover widely used EEG/time-series decoding architectures and recent signal-domain EEG foundation models. We agree that broader comparisons with future EEG foundation models would be valuable. In the camera-ready version, we will more clearly categorize the compared baselines and better position LEGEND against existing EEG foundation models.
Reviewer 3: The two modules may not be individually new; it is unclear which competing methods are foundation models and whether any combine EEG signals with text. A: We thank the reviewer for this constructive comment. We agree that language alignment and flow matching, as general techniques, have been explored in other domains. Our novelty is not in claiming that these two concepts are entirely new in isolation, but in adapting and integrating them for EEG foundation modeling. Specifically, Signal-Language Alignment (SLA) reformulates heterogeneous EEG decoding tasks into an instruction-based generative framework, aligning EEG representations with task-level language semantics and enabling zero-shot inference on unseen datasets. Flow Matching Latent Denoising (FMLD) adapts flow matching to latent-space EEG denoising, encouraging smooth and invariant structural representations instead of reconstructing raw EEG amplitudes. Their combination addresses the two main limitations identified in our paper: insufficient semantic modeling and overreliance on local masked reconstruction. Regarding the competing methods, we will make the categorization clearer in the camera-ready version. Among the compared baselines, EEGNet, EEGConformer, and SPaRCNet are task-specific/end-to-end EEG decoding models, while BIOT, LaBraM, CBraMod, and CSBrain are pretraining-based or foundation-model baselines. These EEG foundation models mainly learn signal-domain representations and typically rely on downstream adaptation. In contrast, LEGEND explicitly aligns EEG signals with natural-language prompts and performs instruction-based answer generation, which distinguishes it from the compared signal-domain EEG foundation models.
Finally, we thank Reviewer 3 for pointing out the reference formatting issues. We will correct the acronyms in the references and add the missing years for the relevant references in the camera-ready version.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
The paper has received 3 positive reviews with all reviewers acknowledging its methodological novelty (language alignment and flow matching) and extensive validation. However all reviewers agree on the need for more details on the method development and a better positioning against the SOTA in EEG analysis. The authors should take care of these concerns.
My recommendation is for the paper to be accepted.
