Abstract

Recent advances in EEG foundation models have predomi nantly relied on self-supervised learning to extract generalizable representations from large-scale datasets. However, this paradigm overlooks the rich semantic information such as subject identities and task labels available in pre-training datasets. Therefore, it limits the model’s ability to consider inter-subject variability and task heterogeneity in EEG datasets, often resulting in suboptimal generalization to unseen subjects and adaptation to various tasks. In this paper, we propose a novel Subject- and Task-Aware EEG Foundation Model (STEM) that fully leverages available semantic information. This method disentangles feature representations into subject-aware and task-aware components by integrating a metric-based meta-learning with contrastive self-supervision. This maximizes the utility of semantic labels to guide the learning of robust features. Extensive experiments across three down- stream benchmarks demonstrate that incorporating semantic supervision significantly enhances adaptation capability for various tasks and new subjects. Our findings suggest that actively utilizing subject and task information during pre-training is crucial for building robust EEG foundation models. The code is available at https://github.com/SionAn/MICCAI2026STEM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2756_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/SionAn/MICCAI2026STEM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{AnSio_Subject_MICCAI2026,
        author = { An, Sion AND Kim, Soopil AND Park, Sang Hyun},
        title = { { Subject- and Task-Aware EEG Foundation Model } },
        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 a novel Subject- and Task-Aware EEG Foundation Model (STEM) that moves beyond conventional self-supervised pretraining by explicitly incorporating semantic information such as subject identities and task labels into the learning process. The framework includes a shared encoder, a decomposer module, and specialized decoders, jointly optimized to learn robust and adaptable representations, leading to improved performance across multiple downstream EEG benchmarks and few-shot adaptation scenarios.

  • 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.

    Introduces a novel paradigm by integrating semantic supervision into EEG foundation models. Effectively disentangles subject-aware and task-aware representations. Combines self-supervised learning with metric-based meta-learning in a unified framework. Addresses key challenges of inter-subject variability and task heterogeneity. Demonstrates strong performance across multiple downstream tasks and datasets. Includes comprehensive evaluation on both task-related and state-related EEG benchmarks. Shows superior efficiency compared to larger foundation models (e. g. , fewer parameters than EEG-DINO). Provides ablation studies validating the contribution of each component. Demonstrates strong few-shot adaptation capability for new subjects. Well-aligned with emerging trends in foundation models for biomedical signals.

  • 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 improvement over strong baselines is relatively modest in some benchmarks (e. g. , Table 2 on page 6). The novelty is partly incremental, as contrastive learning and meta-learning have been previously explored in EEG and representation learning. The disentanglement mechanism lacks deeper theoretical justification or analysis. Limited discussion on computational cost and scalability for large-scale deployment. Generalization across highly diverse real-world datasets is not extensively validated.

  • 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?

    NA

  • 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

    Overall, the approach is promising and aligns well with current trends in foundation model research.

  • 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 paper presents a well-motivated and technically strong contribution by introducing a subject- and task-aware framework for EEG foundation models. The integration of semantic supervision with contrastive learning and meta-learning is both meaningful, addressing key limitations of existing self-supervised approaches. The method is well-designed and supported by comprehensive experiments, including multiple benchmarks, ablation studies, and few-shot adaptation analysis, demonstrating consistent improvements and strong generalization capabilities. While some aspects of the methodology build upon existing techniques and the performance gains are comparable in certain cases, the overall contribution is significant and advances the direction of EEG foundation models.

  • 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 novel framework called STEM for learning generalizable EEG representations. The authors argue that existing EEG self-supervised learning paradigms overlook valuable semantic information such as subject identity and task labels, leading to poor model performance when dealing with inter-subject variability and task heterogeneity. To address this, STEM combines contrastive self-supervised learning with a metric-based meta-learning strategy. The model employs a shared encoder and a Decomposer to disentangle latent features into subject-aware and task-aware components. The authors evaluate STEM on three downstream datasets, demonstrating competitive performance compared to existing baselines.

  • 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 research motivation aligns well with a key point in the field, addressing the problem of existing EEG foundation models neglecting semantic information, and is practically oriented. (2) The model architecture is well-designed and complete, with basic experimental validation carried out, achieving performance improvements on certain downstream datasets.

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

    (1) The feature Decomposer is designed using only simple 1D convolutions, and the theoretical justification for the disentanglement mechanism is insufficient. It is not demonstrated why this architecture can effectively separate subject-specific and task-specific features. The model lacks explicit mathematical constraints to prevent task features from still being entangled with subject information, such as orthogonality loss, adversarial feature learning, or mutual information minimization (2) The related work only surveys self-supervised EEG foundation models, without adequate coverage of related areas such as semantics-aware representation learning, meta-learning-based feature disentanglement, and cross-subject biosignal modeling. The paper fails to clearly differentiate its contributions from existing work on semantics-aware representation learning and cross-subject EEG modeling, resulting in low discriminability of its claimed novelty. (3) The total loss function is a weighted sum of three components, but no justification is provided for the choice of loss weights, and no ablation study is conducted on them. The training stability and reasonableness of this design remain unvalidated. (4) There are no qualitative visualization experiments for representation disentanglement, making it impossible to intuitively verify the actual separation between subject-specific and task-specific features. The validation of the core design is thus insufficient. (5) Since the authors integrate multi-task datasets during pre-training, what is the scaling law of the model after incorporating the proposed feature disentanglement mechanism, and how does performance change accordingly? (6) As a “foundation model,” the experiments must cover a broader range of downstream tasks, such as LOSO emotion recognition or LOSO epilepsy detection. (7) Statistical significance tests are absent from the ablation results tables. The performance gaps between the full STEM model and its ablated variants (w/o Subject and w/o Task) are very small, which further undermines the significance of the proposed feature disentanglement design. (8) Can the authors provide more in-depth empirical analysis to explicitly demonstrate that task representations are fully independent of subject identity and that subject representations are insensitive to task? For example, t-SNE visualizations of features colored separately by subject and by task, or cross-classification accuracy tests.

  • 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?

    Although the motivation of this paper to address EEG signal heterogeneity is clearly articulated, and the work demonstrates solid performance on certain benchmarks, the methodology lacks sufficient depth and constraint mechanisms to substantiate the core claim of “feature disentanglement.” Furthermore, the relatively small model scale and the limited scope of downstream experimental evaluation are insufficient to support the ambitious positioning of this work as a general-purpose “foundation model.” I therefore recommend a weak rejection, and encourage the authors to strengthen the mechanistic design of feature disentanglement and substantially expand the breadth of empirical evaluation in future submissions.

  • 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



Review #3

  • Please describe the contribution of the paper

    The paper proposes STEM, an EEG foundation model that augments self-supervised contrastive pretraining with metric-based meta-learning over subject IDs and task labels. A shared encoder feeds three decoders (self-supervised, subject, task), and a decomposer disentangles latent features into subject-aware and task-aware components. Patch embedding combines a 1D CNN tokenizer with Fourier, 3D-electrode, and sinusoidal position embeddings. Pretraining uses 18 public EEG datasets; evaluation is on PhysioNet-MI, SHU-MI, and ISRUC, including a few-shot adaptation setting.

  • 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.
    • Clear motivation: Using subject and task labels during foundation-model pretraining, rather than treating EEG as unlabeled, is a sensible direction given inter-subject and task variability.
    • The combination of SSL contrastive pretraining with episodic metric-based meta-learning for subject/task disentanglement is not something directly covered by the major EEG foundation models (LaBraM, CBraMod, CSBrain, EEG-DINO), which rely on masked reconstruction or distillation.
    • Parameter-efficient: STEM reports 3.1M parameters vs. up to 201M for EEG-DINO.
    • Few-shot adaptation analysis is a relevant addition given the personalization motivation.
    • Competitive with BIOT and LaBraM-Base on the reported benchmarks.
  • 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.

    Benchmark claims overstated. Under the standard CBraMod/CSBrain evaluation protocol, the published numbers on these same datasets are: PhysioNet-MI BAcc — CBraMod 0.6174–0.6417, CSBrain 0.6304; SHU-MI BAcc — CBraMod 0.6370, CSBrain 0.6417; ISRUC BAcc — CBraMod 0.7865, CSBrain 0.7925.STEM’s numbers (0.6242 / 0.6226 / 0.7863) sit in the BIOT–LaBraM-Base band and are generally below CBraMod and CSBrain, yet the paper frames STEM as SOTA. The authors should clarify whether splits and metrics match the CBraMod protocol, and temper SOTA claims accordingly.

    Novelty of disentanglement is overstated. Subject-invariant / task-relevant EEG disentanglement has substantial prior art not discussed: Özdenizci et al. (IEEE Access 2020, adversarial subject-invariant EEG), Han et al. (arXiv 2008.11426, disentangled adversarial AE) and PTSM (arXiv 2508.11357, 2025, dual-branch task/subject orthogonality). The specific combination with foundation-model pretraining is the real novelty and should be framed as such.

    Missing EEG meta-learning baselines. The central methodological ingredient is meta-learning, yet no EEG meta-learning method is compared — not even the four An et al. works the paper itself cites, nor Ng & Guan 2024 (Neural Networks, subject-as-task meta-learning), SSML (Neural Networks 2023), or MAgML (2025).

    3D electrode-coordinate embedding is not properly positioned. [2510.21585] REVE: A Foundation Model for EEG – Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects already proposes 3D-coordinate-based positional encoding for EEG foundation models and is pretrained at a larger scale; it should be cited and contrasted. Also CoordConformer: Heterogenous EEG datasets decoding using Transformers | OpenReview is another example of coordinate-based positional encoding.

    Anomalous ablation number. In Table 4, “w/ MAE” reports AUC-PR = 0.2204 on SHU-MI, well below the 0.5 chance floor for balanced binary classification and ~0.45 below the other variants. This is almost certainly a reporting error (label inversion or metric bug), since MAE-style EEG SSL (MAEEG, EEGViT-MAE, CBraMod’s own pretext) consistently yields competitive numbers. Please re-verify and additionally report AUROC / BAcc for this cell.

    Undeclared inconsistency. Section 3 mentions “two state-related EEG analysis [14, 19]”, but reference [19] (Mumtaz depression) does not appear in Table 1 or the results.

    EEG-DINO comparison unclear. EEG-DINO does not report PhysioNet-MI / SHU-MI / ISRUC in its MICCAI 2025 paper, so STEM’s EEG-DINO numbers on ISRUC must be from a re-implementation, this should be stated.

    No variance reporting. Gains over the closest baselines are small (e.g., +0.0017 Acc over CBraMod on PhysioNet-MI) with no error bars or significance testing.

    Design choices unjustified. Asymmetric sampling for subjects vs. symmetric for tasks, K=3, and patch length 40 are not ablated.

  • 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?

    The method is clearly described and the motivation to use subject/task semantics in EEG foundation model pretraining is reasonable. However, the contribution as presented has issues: (1) the SOTA claim is not supported when STEM’s numbers are compared against the CBraMod/CSBrain protocol, on which those models outperform STEM on all three target datasets (2) both the disentanglement idea and the 3D-coordinate embedding have direct prior art (Özdenizci, Jeon, Han, PTSM; REVE) that is not engaged with. The core idea of combining SSL pretraining with metric-based meta-learning is worthwhile and should be surfaced more carefully; with clearer benchmarking and reporting of results, better positioning against prior disentanglement / meta-learning work, and corrected ablations.

  • 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 all reviewers and the meta-reviewer for their valuable feedback. We address each concern below, grouped by topic. 1.Novelty (R1.2). While contrastive learning and meta-learning have been used in EEG analysis, existing studies have focused on a single task/dataset setting, without considering generalizability. Moreover, current EEG foundation models solely rely on self-supervised learning without leveraging semantic information. Our contribution lies in leveraging semantic information for EEG foundation model to learn more generalizable EEG representations. 2.Disentanglement (R1.3, R2.1, R2.4, R2.7, R2.8, R3.2). Disentanglement itself is not new in EEG analysis. We adopt the design of [3], but our novelty lies in being the first to integrate disentanglement into an EEG foundation model trained across multiple tasks/datasets via meta-learning, boosting performance and generalizability. For R2’s concerns, all variants in Table 4 show statistically significant improvements (p<0.05). The t-SNE on ISRUC demonstrates that task-dependent features form class-wise clusters across subjects, while subject-dependent features are well separated regardless of class. 3.Comparison (R1.1, R3.1, R3.7, R3.8). We followed the same experimental settings as previous works. Reproducibility issues have been reported on the official baseline GitHub repos; thus, we reproduced baselines with a large gap using official code. EEG-DINO was reproduced as it was not validated on our settings. For R1 and R3’s concerns, the gaps over baselines are statistically significant (p<0.05). We will state this, soften SOTA to consistent improvements, and clarify that some scores come from reproducing with official code. 4.Evaluation scope (R1.5, R2.6, R3.6). We agree that broader validation would further strengthen the work. Notably, we found that STEM also shows similar performance trends on the Mumtaz dataset, supporting the generalizability. 5.Missing baselines (R3.3). Existing EEG meta-learning methods focus on formulating each subject as an individual task to mitigate inter-subject variability, and target single-task few-shot scenarios. In contrast, STEM leverages multiple task labels alongside subject IDs via meta-learning. Due to these different setups, direct comparison is not entirely straightforward. We will clarify these differences in the revised paper. 6.Corrections (R3.5). The correct scores were AUCPR = 0.5231 and AUROC = 0.5827.7.Related work (R2.2, R3.4). We primarily focused on EEG foundation models due to space limitations. We will expand it to cover related studies suggested by R2 and R3.Importantly, while prior works have been conducted on a single task/dataset, our work extends these to a multi-task/dataset setting. 3D-coordinate encoding is not intended as a primary contribution, but rather shares a similar purpose to the related approaches mentioned by R3.8.Additional ablations (R1.4, R2.3, R2.5, R3.9). Pretraining takes ~2 weeks, making ablations such as scalability, loss weights, scaling law, and design choices difficult to complete within the rebuttal period. However, regarding (R2.3) and (R2.5), we verified the effect of loss weights and the scale of training samples during fine-tuning. For (R2.3), task loss weight showed larger variance than subject loss weight, and deviating from the default values led to performance degradation, particularly when weights exceeded 1, as they interfere with the primary objective. For (R2.5), STEM achieves comparable results to the baselines even when fine-tuned with half of the training samples, suggesting the effectiveness of meta-learning in enabling data-efficient fine-tuning. Regarding (R1.4), STEM requires significantly fewer MACs than EEG-DINO, though more than others due to 5× more patches from our smaller patch size. We believe this can be improved through exploration of optimal patch sizes. We agree these analyses would strengthen the work and will discuss them.




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 paper proposes a subject- and task-aware EEG foundation model that integrates semantic supervision with contrastive self-supervised learning and meta-learning. The motivation is well aligned with challenges in EEG representation learning, particularly inter-subject variability and task heterogeneity. The method is evaluated across multiple datasets and includes few-shot adaptation experiments, which are relevant and valuable contributions.

    The reviewers raised several concerns regarding the positioning and validation of the contribution. In particular, the claim of state-of-the-art performance appears overstated, as the reported results are not consistently stronger than existing methods under comparable evaluation protocols. The novelty of the disentanglement mechanism is also partially incremental, with related prior work not sufficiently discussed or differentiated. Also, some inconsistencies in reported results and experimental details further reduce confidence in the claims.

    However, the core idea is promising, and the paper presents a meaningful direction by combining semantic supervision with foundation model pretraining. The concerns raised primarily relate to positioning, validation, and clarity rather than fundamental flaws.

    Given the above, the paper is invited for rebuttal.

    The authors are encouraged to:

    -Clarify benchmarking protocols and moderate claims of state-of-the-art performance. -Strengthen comparison with relevant prior work, particularly in disentanglement and meta-learning. -Include additional baselines or justify their absence. -Provide statistical validation such as variance and significance testing. -Resolve inconsistencies in reported results.

  • 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 rebuttal addressed the major concerns reasonably well by clarifying that the main contribution is the integration of semantic supervision and meta-learning within a multi-task EEG foundation model setting, rather than proposing disentanglement itself as a novel concept. The authors also clarified benchmarking details, corrected inconsistencies in the reported results, and addressed the absence of statistical significance reporting.

    While some limitations remain, particularly regarding broader validation and deeper theoretical analysis, I believe the major concerns were sufficiently addressed and the paper is above the acceptance threshold.



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 sufficiently addressed the major concerns from the reviewers. EEG foundation models can an interesting topic for the MICCAI community.



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 have addressed most of the concerns raised by the reviewers.



back to top