acceptodds
Under review as a conference paper at ICLR 2027

Target-to-Source Alignment: A Hierarchical Geometric Framework with Structural Constraints for Few-Shot EEG Adaptation

Abstract

Few-shot cross-subject EEG seizure prediction faces severe data heterogeneity and optimization instability. While prevalent "symmetric alignment" methods typically seek a compromised shared subspace, they often falter when domain discrepancies are vast, leading to mutual interference and negative transfer. To overcome this, we abandon the pursuit of a middle ground and advocate for a "target-to-source alignment" strategy that explicitly maps target distributions onto the rigid source manifold. To regulate the optimization dynamics of this asymmetric mapping, we present H-GSC (Hierarchical Geometric Framework with Structural Constraints). Specifically, H-GSC employs Feature-Space Guidance (FSG) as a geometric pre-conditioner to reorient adaptation gradients, and Hierarchical Gradient Alignment (H-GA) to filter destructive interference by strictly prioritizing source discriminability. To prevent manifold collapse, we preserve intrinsic manifold structure via a dual-constraint regularization scheme (synergizing Masked Structural Consistency with semantic anchors) and ensure robust early stopping with a Metric-Decoupled Validator. Theoretical analysis confirms that H-GSC aligns with generalization bound minimization. Extensive experiments on CHB-MIT demonstrate state-of-the-art performance (79.54% AUC). Crucially, in rigorous joint cross-dataset scenarios where prior methods suffer from negative transfer, H-GSC achieves a significant 9.65 pp AUC gain, validating that our rectified trajectory effectively bridges vast distributional shifts for scalable clinical deployment.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Few-Shot Domain Adaptation, Seizure Prediction, Geometric Optimization, Feature-Space Guidance, EEG

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kR3WOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigated cross-subject EEG seizure prediction under domain shift and proposes H-GSC, a hierarchical geometric framework with Structural Constraints for few-shot EEG adaptation. I like the motivation, but the experiment is far from solid. First, there is no baseline comparison; the main table is an ablation study. And it only tests on seizure detection tasks; for manifold-based domain adaptation methods for EEG, the community usually tests them on motor imagery, SSVEP, and emotion tasks. I did not see any motivation or design to focus on seizure tasks. I put 1 now; I am open to reconsider if I can see experimental improvements

Critical strengths

S1: Cross-subject decoding is a very related and practical powerful research topic in the field.

S2: The discussion between domain adaptation methods and foundation models for EEG decoding is meaningful.

Critical weaknesses

W1: The main experiment results table is an ablation study.

W2: No comparison with either domain adaptation or EEG decoding baselines in the main contexts. More comparison baselines are needed; the authors discussed these in related work but did not compare with them.

W3: Table 12 violates the ICLR format requirements.

W4: The propsoed framework studied the cross-subject decoding framework in EEG, but only tested on seizure prediction tasks. I did not see any motivation for this framework that is specialized for seizure prediction tasks.

W5: EEG-based manifold decoding is mainly used for motor imagery, SSVEP, and emotion. Very few have applied them to seizure prediction because it largely depends on temporal information. See [1][2][3] for examples.

References: [1] Kobler, Reinmar, et al. "SPD domain-specific batch normalization to crack interpretable unsupervised domain adaptation in EEG." Advances in Neural Information Processing Systems 35 (2022): 6219-6235. [2] Pan, Yue-Ting, Jing-Lun Chou, and Chun-Shu Wei. "MAtt: A manifold attention network for EEG decoding." Advances in Neural Information Processing Systems 35 (2022): 31116-31129.

Questions for authors

Seizure prediction tasks that segement EEG windows may introduce bias in few-shots adaptation setting.

I put a score of 1 now. I am open to reconsidering it if you can add more baselines and more EEG tasks. In addition to the domain adaptation (DA) methods discussed in the related work, [1] is the current SOTA DA framework. For the EEG decoding models, classical models like EEGNet [2] and EEGConformer [3] are necessary; for foundation models, [4] are good.

At least you should also test it on broader EEG tasks like motor imagery and SSVEP.

[1] Li et al. "HEEGNet: Hyperbolic embeddings for EEG." International Conference on Learning Representations. Vol. 2026. 2026. [2] Lawhern, Vernon J., et al. "EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces." Journal of neural engineering 15.5 (2018): 056013. [3] Song, Yonghao, et al. "EEG conformer: Convolutional transformer for EEG decoding and visualization." IEEE Transactions on Neural Systems and Rehabilitation Engineering 31 (2022): 710-719. [4] Jiang, Wei-Bang, Liming Zhao, and Bao-Liang Lu. "Large brain model for learning generic representations with tremendous EEG data in BCI." International Conference on Learning Representations. Vol. 2024. 2024.

Reviewer_VVKFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

H-GSC combines source-directed feature mixing, hierarchical gradient projection, masked consistency, target supervision, and label-free checkpoint selection for cross-patient seizure prediction. Experiments cover CHB-MIT, Siena, and pooled cross-dataset adaptation.

Critical strengths
  1. Relevant adaptation setting. Source-prioritized optimization with a compact backbone is a potentially useful approach to limited-label EEG adaptation.
  2. Useful empirical coverage. Two datasets and component ablations expose performance trade-offs, including weaker and non-monotonic Siena results.
Critical weaknesses
  1. The evaluation needs substantial clarification. Reported AUC repeatedly matches balanced accuracy, raising concern about hard-label rather than continuous-score computation; Table 13 also needs consistent aggregation rules. Appendix C.4 permits one seizure’s pre-ictal sequence to span calibration and testing, without explaining how calibration labels would be available prospectively.
  2. The central guarantees are unsupported. Equation 7 establishes a first-order property, not monotonic source-loss reduction under AdamW or positive target transfer. Appendix A.3 incorrectly uses pointwise domination to bound differences of expectations. Random source pairing is also not shown to satisfy the class-conditional transport assumptions.
Questions for authors
  1. Were AUCs computed from continuous scores? Please provide metric code, consistent aggregation rules, and continuous-score ROC-AUC with uncertainty.
  2. Can you separate complete seizure events across calibration/testing, ensure labels are available before prediction, and report event sensitivity and false alarms per hour?
  3. Can you repair the divergence/transport arguments and distinguish idealized gradient-descent properties from guarantees for the implemented optimizer?
Reviewer_cZ8rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors proposes H-GSC for few-shot cross-subject EEG seizure prediction. It employs Feature-Space Guidance, Hierarchical Gradient Alignment, dual-constraint regularization scheme, and a Metric-Decoupled Validator. The experiments were performed on two public datasets. Overall, the cross-subject seizure prediction is a critical research problem, and the reported results are promising. However, the paper presentation, the theoretical justification and experimental validation require further improvement.

Critical strengths
  1. This paper addresses a challenge research problem in EEG adaptation with substantial cross-subject and cross-dataset distribution shifts. In addition, EEG seizure prediction is an important clinical application.
  2. The target-to-source alignment strategy is well motivated.
  3. Experiments on CHB-MIT, Siena, and joint cross-dataset settings provide initial evidence of potential improvements.
Critical weaknesses
  1. While the research problem is of interest, the methodological novelty appears incremental. The contribution lies in the combination of several components adapted from prior works.
  2. The paper discusses several foundation models such as BIOT, LaBraM, CBraMod, but does not include them in the comparison. What are the advantages of the proposed method over the recent foundation models should be clarified.
  3. The experimental results didn't report statistical significance.
  4. Some experimental details are insufficient such as baseline implementation.
  5. The presentation requires improvement.
Questions for authors

see above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.