acceptodds
Under review as a conference paper at ICLR 2027

Geometry- and Relation-Aware Diffusion for EEG Spatial Super-Resolution

Abstract

Recent electroencephalography (EEG) spatial super-resolution (SR) methods, while showing improved quality by either directly predicting missing signals from visible channels or adapting latent diffusion-based generative modeling, often lack awareness of physiological spatial structure, thereby constraining spatial generation performance. To address this, we reformulate EEG spatial SR as a structured multi-modal conditional generation problem and introduce TopoDiff, a geometry- and relation-aware diffusion framework. Inspired by how human experts interpret spatial EEG patterns, TopoDiff conditions generation on sparse EEG signals together with two complementary structured views: topology-aware image embeddings derived from EEG topographic representations, which provide global geometric context, and dynamic channel-relation graphs, which encode inter-electrode relationships that evolve with temporal dynamics. This design yields a spatially grounded EEG spatial super-resolution framework with consistent performance improvements. Across multiple EEG datasets spanning diverse applications, including SEED/SEED-IV for emotion recognition, PhysioNet motor imagery (MI/MM), and TUSZ for seizure detection, our method achieves substantial gains in generation fidelity and leads to notable improvements in downstream EEG task performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, EEG, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DZjSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper combines global scalp topographic representations, dynamic channel relation graphs, and diffusion models to perform EEG spatial super-resolution. The authors evaluate the approach on multiple EEG datasets across various sparsity ratios and argue that modeling geometric relationships and dynamic electrode dependencies improves reconstruction quality.

Critical strengths

The motivation to use global topographic representations and local channel relation graphs jointly to model the spatial structure of EEG signals is well-founded and reasonable.

The experiments span multiple datasets and include a degree of ablation studies and cross-subject splits.

Critical weaknesses

Methodological innovation is highly limited. Topoplot/visual representations, relation graph modeling, and diffusion reconstruction are all well-established techniques. The paper primarily cascades these existing methods without proposing clear new principles or novel learning objectives. The necessity, synergy, and unique benefits of each specific module—especially when compared to simpler models—are not sufficiently established.

The paper completely fails to test on unseen electrode locations outside the training distribution. The sparse input scenarios simply select a subset from a fixed montage, while the target test electrodes repeatedly appear as supervision targets during training. Consequently, the method only validates channel completion under a fixed layout, which fails to support the broader claims of spatial generalization, geometric generalization, or cross-montage capabilities.

For a method that emphasizes being "geometry-aware," it noticeably lacks fair comparisons with coordinate-queryable or continuous-space models. Furthermore, the practical experiments using real low-density devices primarily report changes in downstream classification tasks, rather than evaluating paired reconstruction fidelity against high-density reference recordings.

If the core topographic encoder relies on an external pre-trained vision model, critical details such as its pre-training source, whether it was frozen or fine-tuned, its computational cost, and its specific functional role must be thoroughly detailed. Without this, it is impossible to determine where the performance gains originate.

Questions for authors

Please incorporate permanent electrode location holdouts: these specific locations must not appear in the training inputs, targets, normalization statistics, or validation selection splits.

Please compare your approach with coordinate-queryable or continuous-field methods using the exact same data splits, coordinates, masking strategies, training budgets, and evaluation metrics.

What is the independent contribution of each proposed module compared to a capacity-matched simple Transformer or a direct diffusion model? Please report comprehensive ablations rather than merely showing the final metrics after removing a component.

In the real low-density device experiments, can you provide paired reconstruction error metrics referenced against synchronized high-density EEG, rather than solely reporting downstream classification performance?

Reviewer_N83KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TopoDiff for EEG spatial super-resolution. It combines topographic image features from a pretrained visual encoder with dynamic graphs built from channel similarities. Both are used to condition a diffusion Transformer for missing-channel reconstruction. Experiments cover four EEG datasets and several downstream tasks.

Critical strengths
  1. Strong empirical results across multiple EEG tasks.

TopoDiff is evaluated on four EEG datasets covering emotion recognition, motor imagery, and seizure detection. It improves reconstruction quality across different super-resolution settings, with larger gains in challenging high-sparsity settings. The improvements also extend to downstream tasks, supporting the practical value of the method.

  1. Useful ablation studies and baseline comparisons.

The dynamic graph outperforms both static graphs and a parameter-matched Transformer. Adding the proposed conditioning modules to ESTFormer also improves reconstruction performance. These results provide evidence that structured conditioning is useful beyond the specific diffusion architecture.

Critical weaknesses
  1. Limited novelty and insufficient evidence for the core contribution.

The main approach combines existing techniques rather than introducing a clearly new modeling principle. Prior EEG studies have already explored electrode geometry and dynamic channel relationships (Tang et al., ICLR 2022; Kotoge et al., NeurIPS 2025), while SRGDiff (Liu et al., ICLR 2026) addresses EEG super-resolution using diffusion. The paper does not clearly distinguish its contribution from these works. More importantly, Table 4 shows that adding topographic conditioning to the graph-based model improves SNR by only 0.026 dB. This raises questions about the necessity of one of the paper's two central components.

  1. Poor presentation and inconsistent methodological and experimental reporting.

The manuscript is poorly organized, with figures and tables frequently appearing far from their discussions or under unrelated appendix sections. The writing is repetitive, and several technical explanations are difficult to follow. More seriously, Section 4.1 and Appendix C.2.2 describe different topographic construction procedures, while Tables 1 and 19 report different results without explaining the evaluation differences. The improvement percentages against SRGDiff in Appendix M also do not match Table 18. These issues make the method difficult to reproduce and reduce confidence in the reported results.

Questions for authors
  1. Novelty over prior work. Previous EEG super-resolution methods have already explored structural and functional channel connectivity (e.g., Tang et al., IEEE T-CYB 2023). What is the key methodological novelty of TopoDiff beyond combining existing spatial modeling and diffusion techniques?

  2. Contribution of topographic conditioning. In Table 4, adding topographic conditioning to the graph-based model improves SNR by only 0.026 dB. Does this component provide meaningful gains on other datasets or at higher SR factors? Please provide the corresponding graph-only results.

  3. Implementation consistency. Section 4.1 and Algorithm 1 describe generating topographic images for each temporal patch, while Appendix C.2.2 describes averaging the entire EEG clip before rendering. Which procedure was actually used to obtain the reported results?

  4. Inconsistent experimental results. Why do Tables 1 and 19 report different reconstruction results? For example, TopoDiff's NMSE on SEED-IV at 4× SR changes from 0.3091 to 0.3402. Are these results based on the same test samples, checkpoints, and evaluation protocol?

  5. Reliability of improvements over ESTFormer. The improvements over ESTFormer on SEED-IV at 2× and 4× SR are very small. However, the five-seed experiments do not include ESTFormer. Do these improvements remain consistent across independent training runs under the same evaluation protocol?

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

This paper proposes a conditional generative framework for EEG spatial super-resolution that reconstructs missing EEG channels from sparse recordings. The framework combines topographic image embeddings for global spatial context, dynamic channel-relation graphs for time-varying inter-electrode relationships, and a Transformer denoiser for signal reconstruction. Experiments evaluate reconstruction fidelity and downstream task performance across emotion recognition, motor-related tasks, and seizure detection using subject-disjoint data splits. Additional analyses examine individual components, alternative reconstruction methods, and robustness under different input conditions.

Critical strengths
  1. TopoDiff combines topographic image features with dynamic channel-relation graphs, providing a clear way to incorporate scalp geometry and changing inter-channel relationships into EEG reconstruction.
  2. Experiments cover emotion recognition, motor-related tasks, and seizure detection, assessing both signal reconstruction quality and downstream task performance.
  3. The appendix includes component ablations, alternative reconstruction baselines, noise robustness tests, and different electrode configurations, helping readers assess the method beyond the main performance tables.
Critical weaknesses

Q1. The results in the main text and appendix are inconsistent. In Table 1, ESTFormer achieves an NMSE of 0.5538 and a PCC of 0.634 on SEED at 2× SR. In Table 19, the corresponding segment-level values are 0.5855 and 0.5694. Please explain whether these results use different models, data, or evaluation procedures, and reconcile the tables. Q2. Section 4.1 and Algorithm 1 describe generating a topographic image for each temporal patch and then averaging the image features. Appendix C.2.2 instead describes averaging the entire EEG clip before generating one image. These procedures provide different inputs to the model. Please identify the actual implementation and correct the inconsistent descriptions. Q3. Appendix C.1 describes left-hand, right-hand, both-hand, and both-feet movement or imagery, but then states that the experiment uses three classes. The manuscript does not explain how these activities are mapped into three classes or whether actual movement and motor imagery are combined. Please provide the exact recording runs and label mapping. Without this information, readers cannot determine what the accuracy values in Table 3 measure. Q4. Tables 6-9 provide five-run results for TopoDiff and RDPI but omit ESTFormer, which is generally the strongest reconstruction baseline in Table 1. Some differences against ESTFormer are small: on SEED-IV at 2×SR, NMSE decreases from 0.2648 to 0.2628. Please repeat ESTFormer under the same protocol and report the uncertainty in the performance difference. Repeated comparisons against RDPI alone do not support a claim of reproducible superiority over the strongest baseline. Q5. Table 14 compares the positional-encoding baseline with the complete TopoDiff model, introducing both the graph module and topographic conditioning. The improvement therefore cannot be attributed to topographic conditioning alone. Table 4 provides a more direct comparison: the graph-only model achieves 3.000 dB, while the complete model achieves 3.026 dB. Please base the discussion on this isolated contribution and establish whether the additional 0.026 dB is reproducible. Q6. Appendix C.2.5 only specifies training and test subject counts. Please identify the validation subjects and explain how checkpoints and graph parameters were selected. Q7. Appendix M states that TopoDiff reduces NMSE by 26.1% compared with SRGDiff under its stated SEED 2× setting. However, Table 18 reports 0.7184 for SRGDiff and 0.4982 for TopoDiff, corresponding to a reduction of approximately 30.65%. Please reconcile this discrepancy and check the other improvement claims against the tables.

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.