acceptodds
Under review as a conference paper at ICLR 2027

Rethinking EEG Spatial Super-Resolution as Dynamic State Decomposition and Filtering

Abstract

EEG spatial super-resolution aims to recover high-density neural signals from sparse electrode observations. Existing methods have achieved substantial progress through spatial modeling, generative reconstruction, and temporal representation learning. However, the current EEG state contains two different sources of information: predictable evolution inherited from preceding states and newly emerging changes revealed by the current observation. Most approaches typically do not model these two components explicitly. In this work, we formulate streaming EEG spatial super-resolution as a sequential state filtering problem and propose Dynamic State Filtering (DSF). Rather than relying on a single temporal reconstruction pathway, DSF combines a history-aware dynamical branch with an innovation-filtering branch that explicitly decomposes waveform-state estimation into prediction and observation-driven correction. To infer how spatially incomplete innovations should affect missing locations, DSF routes the observed innovations across electrodes through a mask-constrained, topology-aware transition operator and incorporates them through uncertainty-aware state correction. The two complementary estimates are finally fused to reconstruct the dense EEG signal. Experiments on SEED, SEED-IV and BCI2000 demonstrate that DSF improves reconstruction fidelity and downstream utility while maintaining robustness under dynamically changing and unreliable sparse observations.

Primary area:
learning on time series and dynamical systems
Keywords:
EEG Spatial Super-resolution, Dynamic State Decomposition

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_h7KeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed a method called Dynamic State Filtering (DSF). It reconstructs dense EEG from sparse streaming recordings through historical state prediction and topology-aware innovation correction, fused with learned channel weights. Tests on SEED, SEED-IV, and BCI2000 include three channel densities, corruption, downstream EEGNet, and frozen EEGPT evaluation.

Critical strengths
  • The proposed causal prediction–correction formulation is appropriate for missing EEG channels.
  • Metrics are computed on unobserved electrodes, and component ablations and fixed downstream classifiers strengthen the evaluation.
Critical weaknesses
  • EEG spatial interpolation cannot create genuinely independent neural information. No evidence is provided in the paper to show that high scalp waveform correlation demonstrates recovery of focal cortical activity or discriminative spatial filtering.
  • The paper evaluates EEG spatial super-resolution primarily using NMSE, PCC, and SNR. While these are appropriate signal-reconstruction metrics, whether they adequately measure the recovery of spatially specific neural information is not justified.
  • The uncertainty gain is explicitly an uncalibrated surrogate, and the filtering branch has at most 20% fusion weight (Appendix C.4); evidence that filtering rather than history-driven reconstruction explains the gains remains incomplete.
  • The manuscript repeatedly refers to a "BCI2000 dataset." Strictly speaking, BCI2000 is a general-purpose BCI software platform, not a dataset. The authors should just use PhysioNet dataset which is commonly used. There are some info that needs to be clarified in the manuscript about the dataset. The original dataset contains 109 participants, sampled at 160 Hz, and uses run-dependent event annotations for motor execution and imagery. The manuscript reports an 88-subject subset, 200-Hz reconstruction settings, and four-class motor-imagery evaluation.
Questions for authors
  1. Why only 88 subjects were used, including the subject identifiers and exclusion criteria.
  2. Whether the original 160 Hz recordings were resampled to 200 Hz, and how this was performed.
  3. Which experimental runs were selected and how the four motor-imagery classes were constructed from the run-dependent annotations.
  4. Whether the reconstruction and downstream classification experiments used identical subject partitions and whether any temporal windows overlapped across partitions.
  5. What is results of the direct comparison between downstream decoding using the original sparse electrodes and decoding using the reconstructed high-density EEG?
3 comments

Dear Reviewer h7Ke, could you please let us know whether our responses, protocol clarifications, and additional experiments address your concerns, and whether you have any remaining or new questions we could clarify?

Thank you again for your careful review. We would appreciate your feedback on our rebuttal and the additional experiments when you have a chance.

We have corrected the PhysioNet dataset description, clarified the preprocessing and subject partitions, and added direct sparse-versus-reconstructed decoding comparisons, regional decoding experiments, and further branch analyses. We also clarify that our results support scalp-signal reconstruction and decoding utility, without claiming recovery of independent neural information or focal cortical sources.

Could you let us know whether these clarifications and additional results address your concerns, or whether any specific issues remain? We would be happy to provide further explanations or evidence.

Thank you again for your time and consideration.

Thank you for answering my questions. I will maintain my score.

Reviewer_sj6cOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Dynamic State Filtering (DSF) for sequential EEG spatial super-resolution. It combines a history-based dynamical branch with topology-aware propagation of observation innovations and learned correction gains. Experiments on three EEG datasets report improvements in missing-channel reconstruction and downstream classification.

Critical strengths
  1. Visibility-constrained innovation routing provides a coherent mechanism for correcting predictions at missing electrodes.
  2. Reconstruction improvements are consistent across the evaluated datasets and sparsity levels, with additional downstream classification results.
Critical weaknesses
  1. The contribution is framed too broadly around prediction–innovation decomposition. The dynamical branch in Eq. (1) follows the same ODE propagation–GRU update structure as the Latent ODE baseline in Eq. (22). Appendix E.1 incorrectly states that the Kalman baseline lacks explicit prediction–innovation decomposition, despite its update in Eq. (21). The advantage therefore cannot be attributed to this distinction; stronger evidence is needed for the specific spatial innovation-routing contribution. Moreover, CAFE, an already cited and directly relevant topology-aware reconstruction method, is missing from the comparisons.

  2. The long-term stability conclusion exceeds the evidence. Figure 4 evaluates a fixed target window using at most 15 windows of context, approximately 60 seconds. This supports stability within the tested range, but does not establish the absence of error accumulation during sustained operation. The claimed role of uncertainty-aware correction in preventing drift is also not isolated through corresponding rollout ablations.

Questions for authors
  1. What evidence distinguishes the proposed innovation-routing mechanism from existing prediction–correction models, and how does DSF compare with CAFE under the same protocol?
  2. Can longer continuous-rollout error curves, including an ablation without uncertainty-aware correction, substantiate the stability claim?
  3. How does DSF perform when an entire contiguous scalp region, such as occipital or central electrodes, is masked throughout the input history, compared with dispersed missing channels at the same visible-channel count?
  4. Under such regional masking, are the SEED emotion-classification and BCI2000 motor-imagery gains retained? Please compare against a classifier trained directly on the retained channels and specify the four BCI2000 classes and runs used.
3 comments

Dear Reviewer sj6c, could you please let us know whether our responses and additional experiments address your concerns, and whether you have any remaining or new questions we could clarify?

Thank you for the additional experiments. I have two remaining questions.

  1. Persistent regional masking. When the entire occipital region is masked, NMSE rises from 0.6390 to 0.9739 on SEED and from 0.1724 to 0.5829 on PhysioNet. How much does historical context help in this setting? Please compare DSF with a model trained and evaluated using only the current window, with the same visible electrodes and scoring targets.
  2. Sparse-trained baseline in Table 4. The Sparse LD inputs are evaluated with a linear classifier trained on HD EEG embeddings. Could you keep EEGPT frozen, train separate linear heads on sparse and DSF-reconstructed embeddings under the same protocol, and report accuracy and macro-F1 for both?

I maintain my Weak Reject recommendation. The main methodological contribution is an EEG-specific innovation-routing design within an established prediction–correction framework. The routing ablation shows a benefit on PhysioNet under clean 4×, but the gains are uneven. On SEED under clean 4×, full DSF reduces NMSE only from 0.3883 to 0.3875 relative to the dynamical branch alone, and errors remain high under persistent regional masking. In my assessment, the demonstrated gains are too limited to establish a substantial advance over existing approaches.

AuthorsOpenReview ↗

Response to Reviewer sj6c: Additional Results on Historical Context and EEGPT Decoding

Thank you for these helpful follow-up questions. DSF uses preceding sparse EEG windows to form a predictive state and routes observation–prediction residuals from visible electrodes to unobserved locations through a state- and topology-conditioned mechanism. Our experiments show that incorporating historical context can improve EEG spatial super-resolution, highlighting the potential of DSF for handling electrode dropout in online brain–computer interfaces.

We provide the requested comparisons on historical context under persistent regional masking and on EEGPT decoding with separately trained linear heads below.

(1) Historical context under persistent occipital masking. We compare full DSF with a current-only reconstruction model trained and evaluated without preceding windows. The entire occipital region remains unobserved throughout all history and target windows, so DSF receives historical observations only from electrodes outside the masked region. Full DSF reduces NMSE by 32.4% on SEED and 30.2% on PhysioNet relative to the current-only model. These results support the importance of historical context for reconstruction even when the entire occipital region remains unobserved throughout the available history.

Both models use the same retained electrodes—14 on SEED and 13 on PhysioNet—and are scored on the same missing occipital targets: 8 on SEED and 5 on PhysioNet. The subject splits, preprocessing, training budget, and evaluated target windows are matched.

DatasetMasked regionCurrent-only NMSE ↓Full DSF NMSE ↓Relative NMSE reduction (%)
SEEDOccipital1.43990.973932.4
PhysioNetOccipital0.83510.582930.2

(2) Frozen EEGPT with separately trained linear heads. The original Table 4 uses a shared linear head trained on ground-truth HD embeddings. We have now added the requested comparison: EEGPT remains frozen, while separate linear heads are trained on sparse EEG embeddings and DSF-reconstructed EEG embeddings. Each head is fitted on the corresponding training-set embeddings and evaluated on the corresponding test-set embeddings. DSF reconstruction improves accuracy by 3.41, 5.83, and 2.94 percentage points on SEED, SEED-IV, and PhysioNet, respectively, with corresponding macro-F1 gains of 3.88, 7.91, and 2.87 points. The improvements persist across all three datasets when each input type has its own separately trained linear head, supporting the practical value of EEG spatial super-resolution for decoding with a frozen EEGPT encoder.

DatasetEEGPT input and linear-head training dataAccuracy (%) ↑Macro-F1 (%) ↑
SEEDSparse EEG39.4237.56
SEEDDSF-reconstructed EEG42.8341.44
SEED-IVSparse EEG29.4525.61
SEED-IVDSF-reconstructed EEG35.2833.52
PhysioNetSparse EEG38.4137.86
PhysioNetDSF-reconstructed EEG41.3540.73

We sincerely thank you for your continued evaluation and would be happy to address any further questions.

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

The authors aim to generate high-density EEG from a sparse set of electrodes based on spatial patterns learned from other high-density EEG and preceding seconds of recording. From the point of view of an EEG researcher, this appears to be EEG upsampling using generative models as opposed to conventional interpolation approaches. The results of this generative model outperform traditional methods and recent generative methods.

Critical strengths
  1. The topic is relevant to neuroscience and related fields. Getting more out of low-density and wearable caps is a practical need.
  2. Physiological sensible choices in data aggregation.
Critical weaknesses
  1. Validation is insufficient. The major contributions of HD EEG are topography and source localization. no source analysis is done while topography comparison is minimal.
  2. key details of EEG preprocessing are lacking. Some choices can leak information from the hidden electrodes.
Questions for authors
  1. Have the authors performed quantitative comparisons of topographies between generated and true HD EEG?
  2. What reference was used?
2 comments
AuthorsOpenReview ↗

We thank Reviewer yWBK for the effort in reviewing our work and providing constructive comments. We address each concern or question below.

W1. Validation is insufficient. The major contributions of HD EEG are topography and source localization. no source analysis is done while topography comparison is minimal.

Q1. Have the authors performed quantitative comparisons of topographies between generated and true HD EEG?

Response. Thank you for this suggestion. To complement the qualitative scalp maps, we have added quantitative comparisons of DSF, CAFE, and spherical spline interpolation on SEED and PhysioNet under the original 4× layout. Spatial PCC measures the Pearson correlation between reconstructed and ground-truth potentials across these electrodes at each time point, assessing agreement in spatial patterns. GFP nRMSE measures the normalized root-mean-square error between their GFP time courses, where GFP is defined as the standard deviation of potentials across the evaluated missing electrodes. It assesses reconstruction of the overall spatial amplitude within this electrode set. DSF achieves higher spatial PCC and lower GFP nRMSE than both comparison methods on both datasets, providing additional evidence for reconstruction of spatial patterns and amplitudes at missing electrodes.

DatasetMethodSpatial PCC ↑GFP nRMSE ↓
SEEDDSF0.4544 ± 0.00090.4287 ± 0.0075
SEEDCAFE0.4074 ± 0.00870.6817 ± 0.0087
SEEDSpherical spline interpolation0.26532.5291
PhysioNetDSF0.7032 ± 0.00060.2404 ± 0.0011
PhysioNetCAFE0.6775 ± 0.00080.2970 ± 0.0098
PhysioNetSpherical spline interpolation0.64730.3323

W2. key details of EEG preprocessing are lacking. Some choices can leak information from the hidden electrodes.

Q2. What reference was used?

Response. Thank you for identifying this issue. Our original preprocessing computed a common-average reference from all electrodes before masking. The reference calculation therefore depended on hidden-electrode signals. To evaluate this factor, we additionally trained DSF using a reference computed only from the 16 visible electrodes under the original 4× layout. DSF performs similarly whether the reference is computed from all electrodes or only the visible ones. These results suggest that DSF’s reconstruction performance is not strongly dependent on including hidden-electrode signals in the reference calculation.

DatasetMetricOriginal all-electrode referenceVisible-electrode mean reference
SEEDNMSE ↓0.38750.3864
SEEDSpatial PCC ↑0.45440.4503
SEEDGFP nRMSE ↓0.42870.4212
PhysioNetNMSE ↓0.13340.1393
PhysioNetSpatial PCC ↑0.70320.7022
PhysioNetGFP nRMSE ↓0.24040.2371

Thank you again for your thoughtful review. We would appreciate your feedback on the additional evidence provided in our rebuttal when you have a chance.

In response to your questions, we have added quantitative topography comparisons using spatial PCC and GFP nRMSE, alongside CAFE and spherical spline interpolation. We have also clarified the original reference procedure and retrained DSF using a reference computed exclusively from visible electrodes, removing hidden-electrode dependence from the reference calculation. We explicitly limit these findings to scalp-signal reconstruction rather than cortical-source recovery.

Could you let us know whether these additions address your concerns about topography validation and preprocessing, or whether further clarification would be helpful?

Thank you again for your time and consideration.

Comments on the paper

2 comments

Dear Area Chair,

Thank you for overseeing our submission #8201. We would appreciate your help in encouraging further discussion. For your convenience, we briefly summarize the main concerns and our responses:

  • Novelty and robustness (sj6c): We clarified that our contribution is state- and topology-conditioned spatial innovation routing, and corrected the Kalman baseline description. Added routing controls support this specific mechanism. We also added CAFE comparisons: CAFE performs better on clean SEED, while DSF performs better on clean PhysioNet and all six tested corruption settings. Further experiments cover continuous rollouts up to 264 seconds with an uncertainty-correction ablation, regional masking, and downstream decoding.

  • Interpretation, filtering utility, and reproducibility (h7Ke): We clarified the actual PhysioNet protocol, including all 109 subjects, preprocessing, and class construction, and verified partition separation. Added branch comparisons and direct sparse-input baselines support reconstruction and decoding utility. Mean decoding gains over sparse inputs are 1.16–2.85 percentage points across the three datasets. We explicitly acknowledge that these results do not establish recovery of independent neural information or focal cortical activity.

  • Topography and referencing (yWBK): We added quantitative spatial PCC and GFP comparisons. We also acknowledged that the original reference used all electrodes and retrained DSF with visible-electrode-only referencing, obtaining similar performance.

We would welcome the reviewers’ assessment of this additional evidence and any remaining concerns that we could clarify during the discussion.

Best regards,
The Authors

AuthorsOpenReview ↗

Overall Response to Reviewers and Summary of Revisions

We sincerely thank Reviewers sj6c, h7Ke, and yWBK for their careful and constructive reviews. We appreciate Reviewer sj6c’s recognition of the coherent visibility-constrained routing mechanism and the reconstruction improvements across the evaluated settings; Reviewer h7Ke’s positive assessment of the missing-electrode evaluation, component ablations, and fixed downstream classifiers; and Reviewer yWBK’s emphasis on the practical relevance of dense EEG reconstruction for low-density and wearable acquisition. Their comments helped us sharpen the technical contribution, strengthen the experimental evidence, and clarify the scope and reproducibility of our findings.

Concretely, we have made the following major clarifications and experimental additions:

  • Sharper identification of the technical contribution. We now emphasize visibility-constrained, state- and topology-conditioned spatial residual routing: how observation residuals available at visible electrodes should inform reconstruction at missing locations. We explicitly acknowledge that prediction–correction is an established principle, clarify the relationship between our dynamical branch and the Latent ODE baseline, and correct the description of the Kalman baseline.

  • Controlled experiments isolating the routing mechanism. We added three alternatives: fixed geometric weights routing residuals, historical-state queries routing current waveforms, and fixed electrode queries routing residuals. Under the tested PhysioNet 4× setting, full routing reduces fused-output NMSE from 0.1904 for the strongest simplified alternative to 0.1334, providing mechanism-specific evidence beyond comparisons of complete models.

  • Additional comparisons with CAFE and challenging observation layouts. We added CAFE comparisons on SEED and PhysioNet under clean observations, noise, channel dropout, and their combination. CAFE performs better on clean SEED, whereas DSF performs better on clean PhysioNet and all six tested corruption settings. We explicitly disclose their different temporal input regimes. We also added persistent central- and occipital-region masking, matched dispersed-missingness controls, and comparisons on common missing regional targets.

  • Continuous-operation evaluation with appropriately bounded conclusions. We distinguish fixed-target reconstruction with varying context from sustained sequential operation. The additional SEED evaluation extends continuous operation to 264 seconds and includes a control using the same frozen model with unit residual-update gain. We report rollout curves and paired early–late errors, while limiting our conclusions to the tested recordings rather than claiming drift-free operation over arbitrary durations.

  • Stronger assessment of spatial fidelity and downstream utility. We added quantitative spatial PCC and GFP comparisons, reconstructed-region decoding, branch-level decoding analyses, and classifiers trained directly on retained sparse electrodes. Under the original 4× layouts, training and testing on DSF reconstructions improves mean accuracy over direct sparse-input classification by 2.85, 1.16, and 2.00 percentage points on SEED, SEED-IV, and PhysioNet, respectively. We distinguish these observed gains from claims of statistical significance, independent neural information, or cortical-source recovery.

  • Correction and disclosure of the PhysioNet protocol. We clarified the use of all 109 participants preprocessing and resampling, and the run-dependent construction of the four motor-imagery classes. We also disclosed the test-subject identities and reported checks for shared subjects, temporal-window identifiers, and identical waveform windows across partitions.

  • Explicit examination of the referencing concern. We acknowledge that the original all-electrode common-average reference depended on hidden-electrode signals. To assess this dependence, we additionally retrained DSF using a reference computed only from the 16 visible electrodes under the original 4× layout. The two reference schemes yielded similar DSF performance on SEED and PhysioNet, providing a targeted sensitivity analysis of the referencing choice.

These additions clarify both the evidence supporting spatial residual routing and the limits of the current evaluation. Detailed responses and supporting results are provided below. We thank the reviewers again and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.