acceptodds
Under review as a conference paper at ICLR 2027

HiMDA: Transferring Hierarchical Microstate Dynamics, Not Labels, for Cross-Task EEG Decoding

Abstract

Cross-task EEG decoding requires transferring knowledge between datasets whose label spaces share no meaningful correspondence, as when operational states must inform emotion recognition or clinical outcome prediction. Conventional feature alignment can conflate acquisition variation with neural temporal organization, whereas transferring a classifier imposes source semantics on an unrelated target task. We propose Hierarchical Microstate Dynamics Adaptation (HiMDA), a label-space-agnostic framework that transfers hierarchical temporal dynamics rather than labels or decision boundaries. HiMDA first constructs a domain-specific observation interface and calibrates polarity-invariant target microstate prototypes under a finite source reference. It then fits Bayesian switching dynamical systems independently in each domain, aligns their slow regimes through label-free dynamics fingerprints, and lets soft regime posteriors condition time-local fast microstate switches. Source slow-regime and regime-conditioned fast-transition operators parameterize finite Dirichlet priors whose posterior means are updated row-wise by target evidence, so well-supported target dynamics depart from the source while sparsely observed transitions remain stabilized. A random-forest readout trained from scratch on the adapted representation defines the target semantics, and no source logit or prediction head is transferred. Under a subject-dependent, trial-level five-fold protocol on SEED-IV, HiMDA achieves accuracy. Under patient-disjoint, leave-one-hospital-out evaluation on I-CARE, it attains an AUROC of and an AUPRC of . Together, these results show that hierarchical dynamics transfer across affective and clinical EEG tasks without source–target class correspondence.

Primary area:
applications to neuroscience & cognitive science
Keywords:
cross-task EEG decoding, EEG microstates, hierarchical temporal dynamics, Bayesian adaptation, label-space-agnostic transfer

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

This paper studies cross-task EEG transfer when the source and target label spaces do not correspond. The core idea is to transfer hierarchical transition dynamics characterized by microstates and BSDS, rather than features or a classification head. Specifically, source prototypes are used to anchor target microstates; dynamics fingerprints align the slow regimes on both sides; the source transition matrices are then used as Dirichlet priors and updated with target data; finally, a random forest is trained on the target data.

Critical strengths

What should be transferred when the label spaces do not correspond is an interesting question in itself; treating transition dynamics as the object of transfer is also a fairly novel idea. However, the current experiments do not yet provide enough evidence to support it.

Critical weaknesses
  1. I cannot understand Table 3. According to Eq. (2) and Eq. (9), λμ and τ affect only the target side. The source prototypes, BSDS, and transition matrices are all estimated separately from source data, so removing the prototype anchor or operator prior should not change the source diagnostic. Yet these are precisely the two ablations that cause the largest drops in source Macro-F1 (26.30 and 14.86 points, respectively). Moreover, the four source classes are themselves created using rules based on the alpha/theta ratio, spectral entropy, and similar measures, while the classifier uses overlapping features. This source metric therefore tells us very little.
  2. I do not think the experiments demonstrate that “dynamics can transfer across tasks.” The differences between target-only and the full model are just 3.36 points on SEED-IV and 0.04 AUROC on I-CARE, with neither variance estimates nor significance tests. A key control is also missing: replacing the prior, at the same τ, with a shuffled prior or one from an unrelated dataset would show whether the benefit comes from the source itself or simply from an additional layer of shrinkage. In addition, CBraMod is compared only with a scratch version; the most direct transfer baseline, fine-tuning a pretrained model, is absent.
Questions for authors
  1. How exactly are the two source columns in Table 3 calculated? How do λμ and τ affect the source diagnostic? Which model produces the source numbers in the target-only row?
  2. Could you add controls using a shuffled prior, a uniform prior, and a prior from another dataset, as well as sensitivity analyses for τA, τT, and λμ? At present, these values are simply specified.
  3. The variances of several baselines in Table 1 are unusual (SVM-DE ±20.05, BiHDM ±14.09), and the SVM-DE result of 56.61% is also considerably lower than figures commonly reported in the literature. How were these baselines tuned?
  4. φ concatenates traditional features such as h_base, but there is no ablation. What are the results using only h_base and only the dynamics features?
  5. What is the justification for using microstate transition structure from healthy people at rest as a prior for the EEG of comatose patients after cardiac arrest? How well do the microstates fit I-CARE data with burst suppression (for example, in terms of GEV)?
Reviewer_zULzOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

HiMDA is a new cross-task EEG transfer method that do not rely on dataset labels and based itself on hierarchical dynamics. The method capture the fast dynamics with microstate and the slow temporal dynamics with Bayesian switching dynamical systems. Several aligment techniques are used to align the source and target dynamics including source-referenced axial prototype adaptation (fast dynamics) or dynamics fingerprints (slow dynamics). The method is evaluared with one source dataset and 2 source datasets (SEED-IV and I-CARE) in subject-dependent or hospital-held-out settings.

Critical strengths
  • The method is orignial, well motivated, and describe with good details.
Critical weaknesses
  • The evaluation is too limited, only one source dataset is used and the transfer is limited to 2 datasets. The method would benefit from an scaling analysis of the source dataset and more competive baslines.
  • The method do not include a discussion or limitation section (e.g., discussing multi source dataset transfer).
Questions for authors

Despite the method's novelty and clear description, I would lean for a rejection of the paper at the current state of the evaluation due to the limited number of datasets tested (1 source and only 2 destinations), missing some key baselines (REVE [1] or LaBraM [2]), and the absence of qualitative results. If the evaluation is strengthened, I am willing to increase the score.

Questions ordered by importance:

  • Q1: Could you explain the motivation of using those specific source and target datasets?
  • Q2: For the source dataset:
    • Did you try to experiment using another source dataset?
    • Did you try to subsample the source dataset to see the effect of the scale of the source dataset on the transfer? What is the relation between source dataset size and transfer performance?
  • Q3: Could you expand the evaluation to additional target datasets? Some recent EEG benchmarks like NeuralBench [3] or EEG-FM-Bench [4] provide standardized protocols and baselines. Those benchmarks could be used to test HiMDA on several additional datasets without needing to reimplement and run baselines.
  • Q4: The method would benefit from adding more competitive baselines, especially recent EEG foundation models for their generalization properties. In particular, some state-of-the-art EEG foundation models like REVE [1] or LaBraM [2] are not included in the baselines.
  • Q5: In Tables 1 and 2:
    • Could you clarify if the method in the protocol-matched results are based on the original method's results or if you reimplemented them?
    • Could you explain why reference-only results are reported if the evaluation does not match with the other results? How should it be interpreted?
    • Could you provide more details about the baselines and the motivation behind the choice of them?
  • Q6: Did you do any qualitative analysis of the method? For example:
    • Did you analyze the alignment or the source and target dataset?
    • As the source-anchored prototypes seem to be a key component of the model, did you find any patterns with them that match physiological results?
  • Q7: Could you discuss the limitations of your model? One interesting question that could be discussed is the adaption to support multiple source datasets, especially if the transfer performance is correlated to the source size.

Additional feedback not affecting the decision assessment:

  • Tables 1 and 2 miss a space between the method name and the citations.
  • Details about the computational time of the method would be a good addition to the supplemental material.

[1] El Ouahidi, Yassine, et al. "REVE: A foundation model for EEG-adapting to any setup with large-scale pretraining on 25,000 subjects." Advances in Neural Information Processing Systems 38 (2026): 22541-22577.

[2] 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.

[3] Banville, Hubert, et al. "NeuralBench: A Unifying Framework to Benchmark NeuroAI Models." arXiv preprint arXiv:2605.08495 (2026).

[4] Xiong, Wei, et al. "EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models." arXiv preprint arXiv:2508.17742 (2025).

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

HiMDA studies whether temporal structure learned from resting-state EEG can improve decoding on tasks with unrelated labels. It aligns polarity-invariant microstate maps across datasets, fits a separate switching dynamical system in each domain, and weights each fast microstate switch by the posterior probability of the slow regime at that time. Source transition probabilities then serve as finite Dirichlet priors that target observations can update. Target labels are used for model selection and to train a new random-forest readout. The paper reports 77.04% trial-level accuracy on subject-dependent SEED-IV and pooled patient-level AUROC/AUPRC of 0.85/0.90 under leave-one-hospital-out evaluation on I-CARE. The corresponding target-only hierarchical model achieves 73.68% and 0.81/0.85.

Critical strengths
  1. The paper tests whether the timing of slow context matters. Equation (7) can distinguish two windows with the same overall fast-switch counts if those switches occur in different slow regimes. The corresponding ablation replaces the time-local posterior with its within-window mean, preserving average regime occupancy. Accuracy falls from 77.04% to 74.38% on SEED-IV, and AUROC/AUPRC falls from 0.85/0.90 to 0.83/0.87 on I-CARE. This comparison gives the proposed coupling a concrete interpretation and supports its predictive value on both targets.

  2. The target-only control makes the transfer result informative. This control retains the hierarchical architecture and target readout while estimating the states and operators entirely from target training data. The full model improves on it by 3.36 percentage points on SEED-IV and by 0.04 AUROC and 0.05 AUPRC on I-CARE. The clinical comparison is particularly valuable because the evaluation holds out an entire hospital at a time and aggregates predictions per patient.

Critical weaknesses

The provenance of three SEED-IV baseline results needs verification. Section 4.1 states that the protocol-matched rows share the same five folds, seeds and trial aggregation. Yet the BiHDM, A-LSTM and SVM (DE) rows match both the means and standard deviations in Li et al.[1], Table I, to two decimal places: 74.35 ± 14.09, 69.50 ± 15.65, and 56.61 ± 20.05, respectively. Li et al. explicitly trained on the first 16 trials and tested on the last eight trials of each participant and session. BiHDM is the strongest external comparator in the matched block, so the reported 2.69-point advantage depends on resolving this discrepancy. Fold-level results would allow the comparison to be checked.

The operator ablation does not isolate the information carried by the source transition structure. Table 3 provides a useful comparison: removing the source operator prior while retaining symmetric smoothing reduces SEED-IV accuracy from 77.04% to 75.69% and I-CARE AUROC/AUPRC from 0.85/0.90 to 0.81/0.88. However, setting (\tau_A=\tau_T=0) removes both the source transition pattern and its additional pseudocounts. The result therefore leaves open whether an equally strong prior with mismatched state correspondence would provide a similar benefit. Table 3 also gives no paired uncertainty for these differences, making it difficult to judge their consistency across participants or hospitals.

The comparison with other methods that could use the same source EEG is incomplete. Table 2 states that CBraMod was trained from scratch, but does not specify the initialization or pretraining data used for EEG2Rep. The target-only ablation shows that access to the resting-state source improves HiMDA. It remains unclear whether transferring explicit transition operators offers an advantage over pretraining an EEG encoder on the same unlabeled recordings and adapting it with the same target supervision.

[1] Li, Y., Wang, L., Zheng, W., Zong, Y., Qi, L., Cui, Z., Zhang, T., & Song, T. (2021). A novel bi-hemispheric discrepancy model for EEG emotion recognition. IEEE Transactions on Cognitive and Developmental Systems, 13(2), 354–367. https://doi.org/10.1109/TCDS.2020.2999337

Questions for authors
  1. Clarification of existing results. Which of the BiHDM, A-LSTM and SVM rows were recomputed on the stated five folds? Could you provide the trial assignments and fold-level results or held-out trial predictions, and specify the unit used to calculate each standard deviation, including HiMDA’s 1.09%? If any rows use published values, how does the comparison change when restricted to verified reruns?
  2. Clarification of the observation model. Which I-CARE release and common channel coordinates did you use in Eq. (1)? The cited v1.0 release provides 18 bipolar channel pairs. How did you map those measurements and the 64-channel source recordings into comparable coordinates for prototype matching? What block duration did you use to estimate slow transitions within each two-second clinical window?
  3. Requested mechanism test. With (\tau_A=20) and (\tau_T=10) held fixed, does the aligned source prior outperform a control that permutes only the source regime identities relative to fixed target states? For SEED-IV, could you report paired accuracy differences across participants? For I-CARE, could you report per-hospital results and confidence intervals for the pooled AUROC and AUPRC differences, obtained by resampling paired patient predictions within each hospital?
  4. Source data available to the baselines. What initialization and pretraining data did you use for EEG2Rep in Table 2? Did any encoder baseline have access to the same unlabeled resting-state recordings as HiMDA and the same labeled I-CARE development folds? If so, what patient-level AUROC and AUPRC did it achieve?
  5. Interpretation of the source-domain ablation. In the source diagnostic in Table 3, which recordings supplied the prior mean in Eq. (9), and which supplied the counts used to update it? If both came from the same smoothed counts, the normalized transition probabilities would be unchanged for any prior concentration. In that case, what accounts for the drop in source macro-F1 from 76.74% to 61.88% when the operator prior is removed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.