Abstract

EEG foundation models pretrained via self-supervised learning promise transferable representations, but their generalization remains limited, especially across diverse clinical datasets. Full fine-tuning is impractical for resource-constrained clinical settings due to high computational requirements. In this work, we investigate whether parameter-efficient self-supervised adaptation, updating only 9–29% of parameters suffices to align representations to target tasks. We evaluate our method on two state-of-the-art models with different pretraining objectives: BIOT (contrastive) and CBraMod (masked reconstruction), and evaluate on three clinical EEG datasets for abnormality detection (TUAB), event classification (TUEV), and seizure detection (CHB-MIT) under both in-distribution and out-of-distribution conditions. SSL adaptation yields consistent gains over linear probing, up to 20x AUCPR. Under a fixed compute budget, peak performance requires only 20–50% of available unlabeled data. Critically, when total window count is fixed, performance remains invariant to patient count, suggesting that performance is dependent on overall temporal window diversity only. Our findings demonstrate that parameter-efficient adaptation enables effective deployment of EEG Foundation models (EEG-FM) with minimal computational overhead and data collection burden. Code available at: https://github.com/c3n-group/efficient-eeg-adapt

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4755_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/c3n-group/efficient-eeg-adapt

Link to the Dataset(s)

TUEV (event classification): https://isip.piconepress.com/projects/nedc/data/tuh_eeg/tuh_eeg_events/v2.0.1/ TUAB (abnormality detection): https://isip.piconepress.com/projects/nedc/data/tuh_eeg/tuh_eeg_abnormal/v3.0.1/ CHB-MIT (seizure detection): https://physionet.org/content/chbmit/1.0.0/

BibTex

@InProceedings{DanMeg_ParameterEfficient_MICCAI2026,
        author = { Dani, Meghal AND Liebe, Stefanie},
        title = { { Parameter-Efficient Self-Supervised Adaptation for EEG-FM under Fixed Computational Budgets } },
        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 parameter-efficient self-supervised learning (SSL) adaptation strategy for EEG foundation models. Through this approach, the authors demonstrate that significant performance improvements can be achieved by updating less than 10% of the model parameters, supported by comprehensive comparative analysis.

  • 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.

    Two EEG foundation models are selected, and a normalized evaluation scheme is introduced to ensure a fair comparison between them; this scheme measures the computational overhead of the optimizer. By adjusting the parameter ratio, the study analyzes how PRAUC varies under different 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 authors propose a parameter-efficient SSL adaptation strategy for EEG foundation models; however, they fail to sufficiently convince the reviewers why this approach must specifically be applied to the EEG domain. While the method may be suitable for a conference like MICCAI due to its use of EEG data, simply analyzing performance with respect to the number of parameters provides little to no novelty. Furthermore, the content in Section 2.1 appears to describe a general evaluation protocol for self-supervised learning rather than a distinct component of the proposed method.

    In other words, it is difficult to determine what specific problem in the MICCAI domain is being effectively addressed. The reviewer’s main concern is that the claimed efficiency in data or parameter usage is not inherently tied to the MICCAI domain and could be evaluated in other domains as well.

  • 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 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

    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?

    see weaknesses

  • 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 #2

  • Please describe the contribution of the paper

    This paper evaluates the adaptation of self-supervised learning (SSL) in the context of foundation models applied to EEG data. The main contributions focus on assessing SSL methods in this specific setting under strict experimental conditions. The paper provides insightful information on foundation models and SSL methods when only limited annotated data are available.

  • 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 experiments are performed under well-controlled conditions.

    The experiment on the trade-off between window size and the number of participants is very interesting.

    The comparison between in-distribution and out-of-distribution settings is insightful.

    The paper is well written and well organized.

  • 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.

    Why is this referred to as zero-shot when the final part of the model is still trained? Even if it is a simple model, these parameters still need to be trained. This aspect is therefore a bit misleading. Moreover, in both cases, the final part of the model is trained. The only difference seems to be whether the backbone is retrained using SSL or not.

    Do you think the results would be similar with a model trained on more data, such as the mentioned LaBraM model?

  • 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

    There are some minor typos:

    • The short title (“Title Suppressed Due to Excessive Length”)
    • “FM” is not defined in the abstract or in the title
    • “CHB-MIT” is sometimes written as “CHBMIT”
  • 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 experiments are well conducted, and the results are insightful for the EEG community.

  • 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.

    I was already convinced that the paper could be published. The experiments are well performed, and I believed that the authors had addressed the reviewers’ comments.



Review #3

  • Please describe the contribution of the paper

    The paper proposes and evaluates a parameter-efficient self-supervised adaptation (SA) strategy for EEG foundation models, in which only the final encoder layer is updated using the model’s original SSL objective on unlabelled target-domain data, before downstream linear probing. The approach is evaluated on two pretrained EEG foundation models with distinct objectives (BIOT, contrastive; CBraMod, masked reconstruction) across three clinical EEG tasks under both in-distribution and out-of-distribution conditions. The authors report improvements over zero-shot linear probing, with the largest gains on OOD tasks. A layer-unfreezing ablation argues that updating only the final layer Pareto-dominates multi-layer and full-parameter updates.

  • 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.
    • Clinically motivated and well-scoped. Parameter-efficient adaptation of EEG foundation models is a genuine need at clinical sites with limited compute and unlabelled data access.

    • Patient-disjoint splits correctly implemented. The authors avoid the common and serious failure mode of subject-level leakage in EEG classification. Official splits are used for TUAB and TUEV. ID/OOD status is explicitly labelled per model–dataset pair.

    • Layer-unfreezing ablation yields a non-obvious finding. Table 2 shows that updating more layers does not improve performance and actively degrades CBraMod. This provides empirical support for the last-layer design choice and is a useful result for the community, contradicting the intuition that more trainable parameters help.

    • Two pretraining paradigms evaluated. Demonstrating the effect on both contrastive (BIOT) and masked-reconstruction (CBraMod) pretraining supports the generality of the finding beyond a single SSL objective.

    • Patients-vs-windows decomposition. Under fixed N = P·W and fixed C, varying P is a thoughtful experimental design that speaks directly to clinical data-collection strategy. The question has been largely ignored in the EEG-FM literature.

    • Reproducibility adequate for a conference paper. Optimiser, learning rates, weight decay, schedule, batch sizes, window lengths, filtering, resampling, channel montage, and splits are all specified. Datasets are public.

  • 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.
    • No supervised baseline. The paper’s framing claims value for self-supervised parameter-efficient adaptation specifically, but the only comparison is against zero-shot linear probing on frozen features. A supervised last-layer fine-tuning control at matched |θ_adapt| on the same labelled target data is not reported. Without it, the reported gains cannot be attributed to the SSL objective. They might simply reflect that updating the last layer on target data (by any means) helps. This is particularly problematic on CHBMIT, where the LP baseline is near-random (BIOT AUCPR 1.53% vs. random baseline classifier 1.48%).

    • The data-efficiency finding is confounded with the fixed-compute design. Under C = N·E held constant, reducing N mechanically increases E (and thus r). The reported conclusion that “20–50% of data is sufficient” is therefore equivalent to “r ≈ 12 is sufficient at this compute level”. This is a repetition-diversity trade-off at one operating point and not a data-efficiency result. In deployment, practitioners would not hold C fixed. They would use the data they have. The finding should either be reframed as a characterisation of the optimal (N, E) trade-off at fixed C, or C should be varied to establish the claim’s robustness.

    • The layer-unfreezing ablation (Table 2) uses a single seed. This is the key ablation justifying the “last-layer” design choice, yet it is run at n = 1 while all other experiments use 5 seeds. The differences reported in Table 2 between L and L,L−1 for BIOT, and between L and All for CBraMod are plausibly within seed noise. “Optimal parameter efficiency” claim is not statistically supported as presented.

    • No statistical testing on main results. Paired significance tests (Wilcoxon signed-rank on paired seed differences, or bootstrap confidence intervals on the difference) would directly support the claimed improvements. Several comparisons have overlapping or near-overlapping SEM bands and the magnitude of statistical evidence cannot be assessed without formal tests.

    • Patient-agnostic claim weakest where it matters clinically. Figure 2 shows that on CHBMIT (the only dataset with a realistically small cohort range) error bars overlap substantially, and no trend test is reported. The authors acknowledge this limitation in the conclusion. The deployment scenario motivating the claim (“longer recordings from smaller cohorts”) is precisely the small-P regime where the evidence is weakest. A formal trend test would materially strengthen this claim.

    • Inconsistent parameter-count. The conclusion states SA updates “<10% of parameters,” but Section 3.2 reports 0.92M of 3.19M = 28.8% for BIOT. The “<10%” claim holds only for CBraMod (9.0%).

  • 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

    Consider reporting the repetition factor r explicitly alongside N in Figure 1, so the repetition-diversity trade-off is visually legible rather than implicit.

    The patient-agnostic finding is consistent with recent SSL results in vision. A brief discussion of whether this generalises from natural images to medical time-series, would enrich the contribution.

    Future work could compare against LoRA, adapters, at matched parameter budgets to place the last-layer SSL adaptation strategy within the broader parameter-efficient fine-tuning landscape.

  • 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 paper addresses a clinically relevant question with several sound elements: patient-disjoint splits, explicit ID/OOD labelling, evaluation across two pretraining objectives, and a layer-unfreezing ablation (Table 2) showing the non-obvious result that updating more layers does not help. Reproducibility is adequate. The paper falls below the acceptance threshold because the central claim, self-supervised parameter-efficient adaptation is valuable, is not isolated by the experimental design. The only baseline is zero-shot linear probing on frozen features. A supervised last-layer fine-tuning control at matched |θ_adapt| on the same target data is absent. This is the one comparison that would establish whether the SSL objective contributes anything beyond “any last-layer update on target data helps.” The deployment motivation (label scarcity) is not instantiated in the experiments (linear probing uses the full labelled training set). The concern is compounded on CHBMIT, where the LP baseline is near-random (AUCPR 1.53% vs. random floor 1.48%). Two secondary issues reinforce this. The “20–50% of data is sufficient” finding (Section 3.3) is mechanically entailed by the fixed-compute design C = N·E. It is a repetition-diversity trade-off at one compute level, not a data-efficiency result. The layer-unfreezing ablation (Table 2) uses a single seed while the rest of the paper uses five. Without the supervised baseline, the central attribution remains unsupported. A supervised comparison provided in the rebuttal, if SA still wins, would move my rating upward.

  • 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 rebuttal addresses the majority of my pre-rebuttal concerns substantively. On statistical testing, the authors provide BCa bootstrap confidence intervals on paired SA−LP deltas across all 5 seeds and 6 model–dataset configurations, with all intervals above zero and large effect sizes (Hedge’s g > 0.8 on 5/6 comparisons). They commit to re-running the layer-unfreezing ablation at 5 seeds, with preliminary results confirming the original single-seed finding. For the patient-count claim, Spearman correlations and linear regression slopes are all non-significant, supporting the patient-agnostic conclusion. The parameter-count inconsistency is acknowledged and will be corrected. The authors explicitly accept my characterisation of the data-efficiency finding as a repetition-diversity trade-off at fixed compute and agree to reframe accordingly. My primary pre-rebuttal concern, absence of a supervised last-layer fine-tuning baseline, is not fully resolved. The authors argue that the LP→SA gain is attributable solely to SA because the linear head is trained identically in both conditions, and that supervised FT does not match their deployment scenario (label scarcity). The first point is correct but does not address whether the SSL objective specifically drives the gain versus any form of last-layer update. The second point is legitimate in principle but is not instantiated in their experiments, since the linear probe uses the full labelled training set. I maintain that a supervised control at matched would strengthen the paper.



Author Feedback

We sincerely thank the reviewers and AC. R2 recommends accept; R1 and R3 weak-reject pending rebuttal. R2 and R3 acknowledge the well-controlled design, ID/OOD comparison with different pretraining objectives, patient-disjoint splits, and patient-window decomposition (R3: largely unaddressed in the EEG-FM literature). We address the AC’s three explicit requests: [A] parameter consistency, [B] supervised-baseline justification, [C] statistical evaluation and, [D] R1’s domain-specificity concern.

[A] Parameter consistency. 10% applies only to CBraMod (0.44M/4.44M=10.0%); BIOT updates 0.92M/3.19M=28.8%. We will revise Abstract/Conclusion with no changes in results necessary.

[B] Justification for omitting supervised fine-tuning (SFT, R3). In Linear Probe (LP) baseline, the pretrained EEG-FM encoder is frozen and only a linear head (LH) is trained on labeled target data. In SSL adaptation (SA), the encoder’s final layer is updated via the model’s own SSL objective on unlabeled target data before the LH is trained. Both train the same LH on the same labeled set with the same encoder frozen at probe time. The LP→SA gain (Table 1) is therefore only attributable to SA, SFT control is not needed to isolate this effect. SFT presumes labeled target-domain data, an assumption that does not match our deployment scenario. Importantly, SSL pretext drives gains across contrastive (BIOT) and masked-reconstruction (CBraMod), indicating the effect lies in the SSL objective, not simply updating extra parameters. Thus, it is a specific instantiation (see Sec 2.1) rather than generic SSL protocol (R1).

[C] Statistical evaluation (R3). (i) Table 1 results: on per-seed SA-LP deltas (5 paired seeds x 6 model-dataset configurations) using bias-corrected and accelerated (BCa) bootstrap confidence intervals (95% CI, 9,999 resamples, B. Efron (1987)) lie above zero when considered for performance metrics(AUCPR for TUAB/CHB-MIT, F1 for TUEV), establishing significance at alpha=0.05.Effect size is large on 5/6 comparisons (Hedge’s g > 0.8). (ii) We will also rerun the layer-unfreezing ablation (Table 2, R3) at 5 seeds; first results confirm that updating only the last layer performs as well as updating two or all layers, with 2-7x fewer parameters. (iii) Patient-agnostic claim (Figure 2, R3): Spearman rank correlation between P and performance is negative in any of the 6 model-dataset pairs (p-value in range 0.24 - 0.97). Similarly, linear regression fits on raw observations per model-dataset pair show that all slope parameters (Figure 2) are not significantly different from 0, suggesting a absence of patient count effect. Full statistical account will be added to the final conclusion.

[D] EEG-specific contribution under clinical-deployment constraints (R1). Our central question is: can we leverage representations of pretrained EEG-FMs and adapt them to OOD clinical settings using unlabeled data — and if so, how much data is needed? We would argue that parameter-efficiency is the deployment constraint. We make 3 specific contributions: (i) bridging OOD shifts in EEG hardware, montage and sampling-rate differences (ii) characterizing the unlabeled target-domain data required for adaptation — we agree with R3 this is a repetition-diversity trade-off at fixed C and will reframe; (iii) showing that longer recordings from few patients are as effective as brief recordings from many.

[E] We acknowledge R2’s typo suggestions (define FM in abstract; standardize CHB-MIT; fix short title); release code on GitHub for reproducibility (R1); rename zero-shot LP to linear probe (LP) throughout per SimCLR/CLIP convention (R2); report repetition factor (r) alongside N in Figure 1 (R3). We will also update our conclusion and future work (R3) to include: SSL LoRA based comparison at matched parameter budgets; generalization to other medical time-series; discussion of whether the patient-agnostic finding transfers from natural images to medical time-series.




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.

    There are inconsistencies in reported parameters, also justify the choice for not using a fully supervised model as baseline. Further statistical evaluations if performed should be reported.

  • 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 in the review.



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 paper addresses an important and clinically relevant problem: parameter-efficient adaptation of EEG foundation models under limited labeled data and constrained computational settings. Reviewers appreciated the patient-disjoint evaluation, explicit ID/OOD analysis, comparisons across multiple SSL paradigms, and the generally strong experimental design and reproducibility. Several concerns were raised regarding the absence of a supervised baseline, the interpretation of the fixed-compute experiments, and the statistical support for some ablations. However, the rebuttal clarified a number of these concerns and appropriately reframed several claims and limitations. While some limitations remain, particularly regarding the SSL-specific attribution, the paper still provides meaningful empirical insights and practical value for the EEG foundation model community. Overall, I support acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This paper provides a nice study on how much data is actually required for adapting current EEG-FM models to achieve reasonable performance when adapted on a new dataset and task. The authors use standard SSL based adaptation and fine-tuning approaches for this and demonstrate their experiments across important benchmarks used in EEG, as well as some ID/OOD settings. There is no particular novelty in their approach, however, it is a systematized study that explores this aspect of EEG-FM.

    I think it’s a valuable contribution to the community of EEG processing, as there have been many EEG-FMs being developed in recent times and this could showcase their capabilities. Although the paper could be improved further in several aspects where more recent FMs could be included, other aspects of validation could be considered, etc. I think it’s a good contribution in it’s current form that could really spark some discussions during the conference.



back to top