List of Papers Browse by Subject Areas Author List
Abstract
Clinical EEG interpretation relies on expert narratives, yet existing self-supervised methods often ignore this textual context. We propose a two-stage multimodal framework using a novel Graph Temporal Relational Encoder (GTRE) that learns montage-agnostic features grounded in 3D brain topology. Stage 1 establishes domain-invariant representations across datasets via multi-domain grounding and prototype-based regularization. Stage 2 (Masked-DPO) introduces a preference-guided objective using expert-validated summary triplets (Fleiss’ kappa=0.71, 90.1% accuracy) to prioritize clinically salient features. By employing a similarity-scaled margin and pathology-aware masking, our model adaptively focuses on diagnostic intent. Results show superior cross-institutional transfer, achieving 96.0% accuracy with only 5% supervision and a +12.7% gain for under-represented seizure types. This highlights the power of preference-based alignment for developing clinically-grounded EEG foundation models.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6477_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AZFAR176/GTRE-
Link to the Dataset(s)
N/A
BibTex
@InProceedings{AzfMoh_GTRE_MICCAI2026,
author = { Azfar, Mohd AND Khan, Izhar Dad},
title = { { GTRE: A Physiological-Semantic Foundation Model for Generalizable EEG Analysis via Preference-Guided Alignment } },
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 multimodal EEG foundation model that integrates physiological signal structure and clinical text through a two-stage alignment framework. The authors introduce a Graph-Temporal Relational Encoder that incorporates spatial brain topology into a transformer architecture to learn montage-agnostic EEG representations. In addition, they propose a two-stage training strategy consisting of (1) multi-domain grounding via contrastive alignment between EEG segments and clinical text, and (2) preference-guided refinement using a masked Direct Preference Optimization objective. The framework aims to improve cross-domain generalization and clinical interpretability, and demonstrates strong performance in downstream EEG classification tasks, including few-shot settings.
- 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.Clear problem formulation addressing cross-domain generalization and the lack of semantic grounding in EEG representation learning. 2.Integration of physiological priors into model architecture, particularly through graph-aware spatial encoding based on electrode topology. 3.A well-structured multimodal learning framework combining EEG signals with clinical text via contrastive and preference-based alignment. 5.Strong empirical evaluation across multiple datasets, including cross-domain transfer, few-shot learning, and rare-class performance.
- 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.Heavy reliance on LLM-generated clinical text and preference pairs, raising concerns about potential bias and noise propagation from synthetic annotations. For example, clinically meaningful uncertainty (e. g. , “possible” or “unclear”) may be removed or altered during LLM-based summarization, leading to overconfident or distorted supervision. As a result, the model may learn to associate EEG patterns with overly certain or stylistically refined descriptions, rather than faithfully capturing the underlying physiological ambiguity. 2.Lack of direct comparison with recent large-scale EEG foundation models under equivalent settings, particularly in terms of model capacity and pretraining scale. While the paper presents comparisons with multiple foundation models, it is unclear whether the experimental conditions are properly matched. In particular, differences in model capacity (e. g. , parameter count), pretraining scale, and fine-tuning protocols are not clearly controlled or reported. 3.Potential over-reliance on textual supervision, which may limit applicability in scenarios where high-quality clinical reports are unavailable or noisy. 4.Insufficient discussion of computational complexity and scalability, especially given the multi-stage training and multimodal alignment pipeline.
- 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?
This paper proposes a multimodal EEG representation learning framework that integrates physiological structure and clinical semantics through a two-stage alignment strategy. The approach is well-motivated and demonstrates strong empirical performance across multiple datasets and evaluation settings. While the individual components are largely based on existing techniques, the overall integration and application to EEG–text alignment is meaningful and practically relevant. The preference-guided refinement stage is particularly interesting, although its novelty is incremental. Some limitations remain, including reliance on LLM-generated supervision and limited analysis of component-wise contributions. Nevertheless, the combination of solid experimental results and a well-structured framework makes this paper a reasonable candidate for acceptance.
- 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.
All my concerns have been addressed in the manuscript and the response letter.
Review #2
- Please describe the contribution of the paper
The authors propose a two-stage alignment framework between EEG signals and clinical reports. In the first stage, they aim to align EEGs collected from multiple centers with clinical entities extracted from the reports. In the second stage, they optimize a policy to rank the reports. These pretraining stages prove effective for downstream seizure and abnormality classification tasks, outperforming state-of-the-art methods. The experimental validation supports the authors’ design choices. The main contribution is the proposal of a smart way to pretrain and align reports and EEGs.
- 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.
- Although the idea is methodologically compositional with respect to existing techniques, it is well structured and each component is justified.
- The experimental validation is complete in its essential aspects and reflects a solid research methodology.
- Results clearly prove the effectiveness of the method.
- 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 presentation of the paper is poor. The narrative coherence, the highly dense language, and the quality of the figures (which, in my opinion, should be totally revised) and tables are not at all on par with the quality of the idea and method. This perhaps suggests that the authors were short on time.
- The related work section is really limited. This is a weakness that should be acknowledged, even though I understand the constraints imposed by the page limit.
-
Some methodological aspects are critical and insufficiently clear. Specifically:
- It is not clear whether the validation of the reports refers to the fact that they remain valid and understandable after summarization, or whether they retain clinically valid meaning even if not correlated with the EEG. In other words, how are they validated? What exactly were the physicians asked?
- In the dataset section, the paper mentions the curation of a set of tokens with specific terms to support alignment with the prototypes, whereas later it states that the prototypes are learnable. I assume this is just an imprecise use of terminology, but it creates confusion.
- It is not clear how the temporal EEG features are produced. From the initial description, it seems that temporal information is compressed into a single token per channel. Then a temporal hierarchy is mentioned, but it is only partially illustrated in the figure, which is difficult to understand. In particular, (Z_{micro}) and (Z_{macro}) are not mentioned in the text. Therefore, I do not understand whether the token preserves temporal information, and to what extent. Likewise, when the output is divided into temporal segments, it is not clear according to which criterion. I believe this issue could be addressed by devoting more effort to designing a clearer figure, revising the notation, and adding the dimensionality of each variable.
- It is not clear how downstream classification is performed. The evaluation part is left implicit. Is a classification head trained? If so, which one? Is the evaluation done in a zero-shot fashion? How?
- The evaluation reports many experiments, but the commentary is often confusing and the meaning of the individual experiments is not always immediately clear. The ablation section is missing some components used in the architecture whose contribution remains unclear, such as location embeddings and geometric bias, even though space is devoted to them in the method section.
- Given the impossibility of including all necessary details due to space limitations, and therefore the difficulty of reproducing the results, releasing the code and model weights becomes essential for the impact of this research.
- Please rate the clarity and organization of this paper
Poor
- 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 provide sufficient information for 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
Figure 3 contains tables and the image is not readable. Moreover, part (e) is never referenced and is impossible to interpret.
- 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?
Strong idea and promising empirical results, but the current manuscript suffers from substantial clarity, presentation, and reproducibility issues that make the method difficult to fully assess.
- 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
Authors propose GTRE, a multimodal foundation model for EEG analysis. It addresses electrode heterogeneity through a Graph-Temporal Relational Encoder that uses 3D topographical coordinate embeddings. A two-stage training process is introduced: Stage 1 involves fine-grained segmental alignment with clinical reports, and Stage 2 utilizes “Masked-DPO,” a preference-guided alignment strategy that uses expert-validated summary pairs to focus the model on clinically relevant diagnostic features rather than lexical boilerplate. The model is validated across multiple clinical corpora (TUSZ, TUAB, MIMIC-IV).
- 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.
- Novel Task Formulation: The application of DPO to the alignment of medical signals and reports is a creative attempt to utilize trendy NLP alignment techniques in a specialized clinical domain.
- Addressing Montage Heterogeneity: The use of 3D coordinate embeddings is a well-motivated response to the persistent problem of varying electrode configurations in EEG research, providing a framework for cross-dataset scaling.
- Large-Scale Data Integration: The effort to combine and standardize three major clinical EEG corpora (TUSZ, TUAB, MIMIC-IV) for pretraining is commendable and provides a high-volume dataset for representation learning.
- Empirical Transfer Results: The reported zero-shot performance on unseen hospital data is a strong indicator of the model’s potential for institutional generalization, provided the datasets are sufficiently diverse.
- 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.
- Hypothesized vs. Validated Priors (Geometric Bias): The distance-based penalty enforces a strictly local interaction prior. However, EEG signals are heavily influenced by volume conduction and global rhythmic synchrony (e.g., generalized seizures). By forcing the model to prioritize local features, the authors may be logically “blinding” the model to long-range functional coordination. There is no evidence provided that this “physiologically-grounded” bias actually outperforms a standard, unconstrained Transformer attention mechanism.
- Over-Engineering (Masked-DPO): While innovative, the use of DPO for filtering clinical text feels like a “sledgehammer to crack a nut.” The authors do not compare this complex, multi-stage preference-learning pipeline against simpler and more standard alignment strategies, such as Hard Negative Mining or basic Entity-Based Masking in a standard CLIP-style objective. Without these baselines, it is unclear if the complexity of DPO is justified. DPO is mathematically derived for generative models. In this encoder-only setting, the authors substitute probability with dot-product similarity. This lacks theoretical rigor; unnormalized similarities do not behave like the log-likelihoods required by the DPO derivation, making the optimization objective’s convergence and meaning questionable.
- Visualization and Transparency: The extremely small font sizes in Figures 2 and 3 make the complex architecture and results nearly illegible, hindering the community’s ability to scrutinize the multi-stage pipeline.
- 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?
GTRE is an ambitious and high-impact submission that introduces a multimodal alignment framework for EEG analysis. While I have identified several theoretical and logical “cracks”—specifically regarding the mathematical validity of DPO for encoder-only models and the potential suppression of long-range brain connectivity—the empirical results are too significant to ignore. Achieving 91.5% zero-shot transfer across institutions and a +12% gain on rare seizure types provides strong evidence that the learned representations are robust and clinically relevant. Although the pipeline is complex and potentially over-engineered, the task formulation is highly original and addresses a critical bottleneck in EEG research. Therefore, I believe the paper’s empirical contributions and methodological novelty place it marginally above the threshold for MICCAI, provided the authors address the theoretical gaps in their rebuttal.
- 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 thank all the reviewers for their feedback, acknowledge concerns on density and legibility, and commit to revising the Introduction and Related Work for clarity.
[MR, R2, R5] LLM Summaries and Clinical Validity. Physicians were given structured rubrics:(1) verify whether source clinical meaning from Findings/Impression was faithfully preserved, and (2) judge whether t+ was more diagnostically informative than t-. Inter-rater agreement reached Fleiss’ \kappa=0.71; 90.1% of summaries and 87.9% of preference pairs were validated by experts. A preferred summary retains seizure type and context (e.g., GNSZ in an unresponsive ICU patient on Ativan/Propofol) while the non-preferred version omits both. Mask-DPO restricts gradients to pathology-aware masked entities. Textual supervision is strictly offline; at inference, GTRE is a pure EEG encoder requiring no clinical reports.
[MR, R2, R5] Comparisons, Capacity, Scalability, and Reproducibility. Table 2 compares GTRE against LaBraM (2024) and FEMBA (2025) at matched capacity (approx 32M params) and pretraining volume; GTRE’s gain (94.3% vs. 93.0%) is architectural. Complexity scales linearly with EEG segments; test-time overhead is approx14 ms/segment on an NVIDIA H100.Fig. 3(b) ablates prototype regularization, margin, and raw reports; Fig. 3(e) benchmarks robustness. We will add a parameter/complexity table and release code, weights, and de-identified examples upon acceptance.
[MR, R5] Temporal Compression, Dimensions, and Evaluation. Each Zmicro(200ms) patch preserves local morphological transients (e.g., spikes) via positional encoding; Zmacro(5s,d=256) segments aggregate context via self-attention, preserving rhythmic flow. Backbone: L=8 layers, H=8 heads, single-layer MLP head. Few-shot (5% labels) freezes the encoder; SOTA comparisons use full fine-tuning on subject disjoint splits. Zeroshot transfer assigns classes via cosine similarity to medical entity embeddings, requiring no task-specific training.
[MR, R2, R5, R6] Clarity and Presentation. Curated token sets are clinical entity vocabularies for masking/alignment, not fixed prototypes; Stage 1/2 prototypes are learnable latent vectors. We will clarify classifier protocol, segmentation criterion, dimensions, terminology, and flow, and rewrite the Introduction and Related Work for clarity. Figs 2–3 will be redrawn with larger fonts; embedded tables moved out as standalone figures; Fig. 3(e) redrawn for legibility.
[MR, R6] Geometric Bias and Volume Conduction. The geometric term is a learnable soft bias regularizing montage heterogeneity and noise—full QK^{\top} attention is preserved, so distant electrodes can attend freely. We agree volume conduction creates genuine long-range correlations; the geometric prior does not suppress these as it is content-agnostic and overridden by learned weights. Fig. 3(e): GTRE holds 91.2% accuracy under extreme sensor dropout vs. 85.2% for standard transformers.
[MR, R6] Mask-DPO vs. Simpler Baselines. HNM fails structurally: boilerplate-heavy clinical reports allow contrastive models to minimize loss via surface phrasing, not pathology-discriminative features. Uniform masking offers no remedy as randomly masked tokens are clinically uninformative. CLIP-style objectives cannot distinguish t+, t- when both describe the same recording, and the text→EEG direction adds noise. Mask-DPO fixes each: it ranks t+ over t- on diagnostic entities only, concentrates gradients via a clinical vocabulary, and is unidirectional. Without Stage 1, Bradley-Terry scores are meaningless and Mask-DPO diverges. Fig. 3(b): HNM drops 11.3%; uniform masking 3.1%; CLIP-style 8.7%. Eq. 3 uses dot-product similarities, not generative likelihoods.
[MR] Synthesis. By grounding morphological transients Zmicro within rhythmic context Zmacro and aligning both against validated clinical narratives via staged preference optimization, GTRE learns montage-agnostic representations without clinical reports at inference.
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.
This work proposes a two-stage alignment framework between EEG signals and clinical reports. Specifically, the authors name it as Graph-Temporal Relational Encoder (GTRE) that incorporates spatial brain topology into a transformer architecture to learn montage-agnostic EEG representations. Post that, the two-stage training strategy comprises (1) multi-domain grounding via contrastive alignment between EEG segments and clinical text, and (2) preference-guided refinement using a masked Direct Preference Optimization objective.
All reviewers have recognized and appreciated the motivation behind this work, especially as there are several heterogenous EEG datasets. The reviewers have also appreciated the smart way of utilizing masked DPO in the framework. The reviewers have also noticed the strong empirical results across the datasets, especially the reported zero-shot performance on unseen hospital data is a strong justification for the framework.
However, there have been equally important concerns raised by the reviewers, particularly, in relation to the writing and presentation (especially the figures), as well as, the overall understanding of the methodology. Please focus on the below points during the rebuttal. I suggest the authors provide clear and reasonable responses to improve presentation and clarity, as significant revisions to the paper are not feasible.
1.Visualization, Transparency, Clarity, and Presenation (R#5,R#6): The extremely small font sizes in Figures 2 and 3 make the complex architecture and results nearly illegible, hindering the community’s ability to scrutinize the multi-stage pipeline. The narrative coherence, the highly dense language, especially in the Introduction and Related Work sections is really poor to understand.
2.Validation of EEG Reports (R#2, R#5): How were the LLM-generated reports validated? What exactly were the physicians asked? How should the quality be trusted?
3.Temporal information compression (R#5): Whether the token preserves temporal information, and to what extent?
4.Clarification on downstream classification (R#5): Was it fine-tuning of a MLP head? How was zero-shot performed?
5.Missing ablations and discussions in results (R#5,R#6): Some components contribution and discussion on experiments is not clear. Specifically, on the geometric bias is introduced by the distance-based penalty, which enforces a strictly local interaction prior, and EEG signals are heavily influenced by volume conduction.
6.Lack of discussion on scalability, model size and parameters (R#2).
7.Masked-DPO Justification (R#6): The authors do not compare this complex, multi-stage preference-learning pipeline against simpler and more standard alignment strategies, such as Hard Negative Mining or basic Entity-Based Masking in a standard CLIP-style objective.
- 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.
This work has introduced a new framework called Graph-Temporal Relational Encoder (GTRE) that incorporates spatial brain topology into a transformer architecture to learn montage-agnostic EEG representations. While the idea is interesting and acknowledged by the reviewers during the first round, there were major concerns raised regarding the presentation, quality of figures, claims, overall understanding of the methodology.
The authors have tried to address several comments during the rebuttal especially relating to the validation of the EEG reports and the role of the physicians in it, temporal EEG features generation, and zero-shot performance. I appreciate the authors efforts to clarify many of such important issues.
I appreciate the authors efforts during their rebuttal. However, the manuscript will require significant revision in presentation, writing, methodology description and figures to ensure that the eventual readers are able to understand this paper in a reasonable manner. Thus, it is hard to provide an acceptance at this stage.
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 reviewers raised concerns regarding the clarity and presentation of the manuscript, validation of the LLM-generated clinical reports, justification of the preference-guided alignment strategy, and several implementation details related to downstream evaluation and scalability. The rebuttal addressed these concerns satisfactorily by providing additional details on the physician validation protocol, report quality assessment, model capacity, computational complexity, downstream classification procedure, zero-shot evaluation strategy, and the rationale behind the proposed Mask-DPO framework.
The paper presents a technically ambitious and well-motivated multimodal foundation model for EEG analysis, combining physiological priors, EEG-text alignment, and preference-guided refinement. The empirical evaluation is comprehensive, including cross-domain transfer, few-shot learning, rare-class recognition, and zero-shot evaluation on unseen hospital data. While concerns regarding manuscript clarity and figure readability remain valid, these primarily affect presentation rather than the technical contribution itself.
The rebuttal successfully addressed the major concerns raised during review and that the paper is above the acceptance threshold.
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.
The authors addressed several concerns, it is recommended to improve the presentation and clarity of methods and results in the revised paper.