acceptodds
Under review as a conference paper at ICLR 2027

Spatial Encoding Geometry Determines the Maps and Transitions Recovered by Deep EEG Microstate Models

Abstract

Unsupervised deep clustering is widely used for discovery in neuroscience, yet model choice is treated as an implementation detail. Models sharing the same data and evaluation metrics are assumed to recover equivalent structure. This assumption is tested in EEG microstate analysis, where discovered states feed biomarker studies. Two variational-deep-embedding models are compared on 203 paired eyes-closed recordings (LEMON). The comparison is centred on how the topography is spatially encoded: one model interpolates scalp topographies onto a 2-D image grid, the other learns on the anatomical electrode graph, with shared preprocessing, peak extraction, and evaluation throughout. As deployed, the two pipelines report a two-fold difference in temporal parameters: mean dwell 70.2 against 143.2 ms and occurrence 3.57 against 1.65 s⁻¹. Decoding each model's own recovered topographies with a common backfit removes that difference: mean dwell 70.2 against 71.0 ms and occurrence 3.57 against 3.37 s⁻¹. The reported divergence is therefore a property of the decode rather than of the encoding. What the encoding does change is the recovered maps. Under the common decode the graph model attains the higher backfit global explained variance, 0.618 against 0.541. Internal cluster metrics do not track this: the image model attains the higher silhouette (0.175 against 0.170, significant after Holm correction), and the indices are computed in different input geometries. These findings indicate that spatial encoding changes the recovered maps, the latent geometry, and the transition statistics of the sequences, but not the dwell and occurrence of the states, and that reported microstate parameters should state the backfit and the decode that produced them.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
EEG microstates, deep generative clustering, graph neural networks, unsupervised learning, global explained variance, variational autoencoders, electroencephalography (EEG), model evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hLJtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper compares two EEG microstate discovery pipelines: CNN-VaDE, which interpolates 61-channel scalp potential topographies into 40×40 images, and Graph-VaDE, which performs convolutions directly on the electrode layout. Experiments were conducted on 203 closed-eye resting-state EEG recordings from the LEMON dataset, using consistent input preprocessing and GFP-peak selection. The two methods were evaluated via subject-matched analyses—comparing spatial maps, Global Explained Variance (GEV), mean duration, occurrence, and transition/complexity statistics—with a modified K-means algorithm included as a benchmark.

Critical strengths
  1. For the same participants and identical inputs, two native pipelines exhibit an approximately twofold difference in "dwell" duration; however, this discrepancy largely disappears following a unified decoding process. This finding has direct implications for cross-method comparisons and holds greater scientific value than the mere pursuit of higher clustering scores.

  2. Based on 203 paired subjects, we examined the GEV for K=3–20K=3–20. Our primary analysis employed the Wilcoxon test, Holm correction, and bootstrap confidence intervals, while also incorporating a modified K-means approach. For K=4K=4, the mean GEV is 0.625, slightly higher than the 0.618 obtained via the graphical model.

Critical weaknesses
  1. The core causal claim—that "geometry determines mapping and transfer"—is confounded. Beyond differences in input geometry, graph-based models and CNNs also diverge in terms of latent variable dimensionality and loss function components. While a unified decoding framework mitigates temporal-statistical confounds, it fails to isolate whether topological differences stem from input representation, model capacity, optimization, or the loss function. Consequently, attributing the observed results to "spatial encoding" remains insufficiently supported.

  2. Limited validation regarding extrapolation and independence. The study relies on a single dataset, a single state condition, and a single montage. Furthermore, because each model is trained on the GFP peaks of the very recordings used for subsequent reconstruction, the reported GEV (Global Explained Variance) reflects in-sample fitting rather than generalization to an independent test set. The authors also acknowledge that the same cohort of 10 participants was used for training-seed selection, peak-level random partitioning, and model architecture search. Therefore, the finding that "graph-based methods yield higher GEV" does not necessarily imply greater reliability on new, unseen recordings; the study lacks assessments of seed stability and test–retest reliability.

Questions for authors
  1. Under identical conditions—including latent variable dimensionality, loss weights, parameter budget, and optimization settings—do the findings regarding GEV, map correspondence, and transition matrices (derived via a unified decoding approach) still hold when varying the 2D interpolation method or the electrode-map representation?

  2. When the training/test split is based on individual subjects—evaluating the model on recordings not used during the fitting process—do the observed advantages in GEV and map stability persist? Could you provide results across multiple random seeds and test–retest intervals?

  3. How can the reported advantages in transition statistics coexist with the observation that dwell times and occurrence rates are essentially aligned? Please include sensitivity analyses regarding the smoothing kernel, minimum segment duration, microstate matching/permutation procedures, and the selection of KK.

Reviewer_ESGdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper compares image-based and graph-based VaDE pipelines for EEG microstate discovery on 203 LEMON recordings. A common decoding procedure removes their large differences in dwell time and occurrence, while differences in recovered maps and transition statistics remain.

Critical strengths
  • The paired analysis across 203 participants provides a useful reminder that decoding and smoothing settings can substantially affect reported microstate parameters.
Critical weaknesses

The evidence remains limited to a preliminary case study:

  • Confounded comparison. The pipelines differ not only in spatial encoding but also in latent dimension, loss terms, and decoder architecture (Table 1). Aligning the final decoding procedure does not remove these differences. The remaining effects therefore cannot be attributed specifically to spatial encoding geometry, as the title suggests.
  • Limited evaluation. As an empirical study, the paper requires validation across multiple datasets, including larger-scale cohorts. The single-dataset evaluation is insufficient to establish the claim.

Minor

  • Figure 1 is barely readable.
Questions for authors

See weaknesses.

Reviewer_aEfhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This study compares image-based and graph-based deep EEG microstate models using 203 eyes-closed recordings from the LEMON dataset. The evaluation covers recovered topographies, explained variance, temporal parameters, clustering quality, and sequence dynamics, with modified K-means included as a classical reference. A useful finding is that applying a common decoding procedure largely removes the differences in dwell time and occurrence between the two deep models, while differences in recovered maps and transition statistics remain.

Critical strengths
  1. Applying a common decoding procedure shows that much of the apparent difference in temporal parameters comes from processing choices, helping prevent misleading comparisons between models.
  2. Comparing both models on the same participants reduces differences caused by participant selection, while experiments across multiple numbers of microstates broaden the analysis.
  3. The study examines map quality, temporal parameters, and sequence dynamics, and includes modified K-means as a classical reference.
Critical weaknesses

Q1. Common decoding substantially reduces the temporal differences, but does not establish that spatial encoding has no effect. Occurrence remains 3.57 versus 3.37 events per second, a difference of approximately 6%. Please define and justify an acceptable equivalence range, or revise the conclusion to state that common decoding reduces, rather than eliminates, the differences. Q2. The original CNN settings responsible for approximately doubling dwell time are not clearly reported. Please specify the original settings and show which change produces the difference. Q3. The discussion compares a graph-model value of 0.728 across 203 participants with a CNN cohort summary of 0.400 across 196 participants. Different participant sets and summary procedures cannot establish a fair advantage. Please calculate template agreement for both models using the same participants, templates, and matching procedure, and report a paired comparison. Q4. The comparison changes several factors simultaneously, including spatial encoding, latent dimension, decoder architecture, and loss terms. Therefore, the observed differences cannot be attributed specifically to spatial encoding. Please isolate these factors experimentally or revise the title and conclusions to describe differences between complete pipelines. Q5. Cluster-validity scores are calculated in different spaces: 1,600 image pixels for the CNN and 61 electrode values for the graph model. Their numerical ordering therefore does not establish a fair performance ranking. Please evaluate both cluster assignments in the same space before interpreting score differences or their statistical significance. Q6. The caption of Table 2 incorrectly states that the graph model has a longer dwell, although the reported values show the opposite. Q7. The labels in Figure 1 and the text in Table 4 should be enlarged for readability.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.