acceptodds
Under review as a conference paper at ICLR 2027

GLAFE: Whole-Brain EEG-to-FMRI Prediction via Global–Local Modeling

Abstract

Functional magnetic resonance imaging (fMRI) provides spatially detailed measurements of whole-brain activity, but its high cost and limited accessibility constrain its use. In contrast, electroencephalography (EEG) is inexpensive, portable, and temporally precise, yet offers limited spatial resolution. Whole-brain EEG-to-fMRI translation could thus enable accessible and detailed mapping of brain activity. Such translation requires both temporal and spatial fidelity: reconstructing the temporal dynamics of individual brain regions of interest (ROIs), while preserving the spatial structure of correlations between them, as captured by their functional connectivity (FC). fMRI signals are composed of a global component, shared across all brain regions, and region-specific fluctuations which deviate from it. Motivated by this structure, we introduce , a lobal-ocal utoregressive framework for predicting whole-brain MRI from EG. By modeling these components separately, our approach is able to reconstruct temporal dynamics while preserving accurate FC structure. To predict the local signals, GLAFE learns region-specific embeddings that query EEG features through cross-attention, while an autoregressive component uses previously generated whole-brain fMRI to refine these predictions. We evaluate our method across multiple datasets and subject splits, and demonstrate its generalization capabilities to unseen datasets and tasks in a zero-shot setting. Across these evaluations, GLAFE outperforms competitive baselines in both temporal dynamics and FC reconstruction metrics in almost all configurations.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuroimaging, EEG, fMRI, Multimodal learning, Neuroscience

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QqzpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whole-brain fMRI prediction from EEG using a method called GLAFE. The main idea is to separate BOLD signals into global and local components. The authors use pretrained REVE to extract EEG features and ROI-specific cross-attention to predict local signals. They also use BrainLM for autoregressive correction based on previous fMRI predictions. The method is evaluated on three simultaneous EEG-fMRI datasets, covering resting-state, auditory-task, and neurofeedback data. The results show that GLAFE generally outperforms several baselines in BOLD prediction and functional connectivity reconstruction. The authors also provide ablation studies and attention map visualizations to analyze different components of the method. Overall, the paper combines global-local modeling, ROI-specific attention, and autoregressive correction for EEG-to-fMRI prediction, with experiments across different datasets and settings.

Critical strengths
  1. The global-local decomposition provides a reasonable way to separately model shared brain activity and ROI-specific fluctuations. I think this is an interesting design for EEG-to-fMRI prediction, especially when both temporal dynamics and functional connectivity need to be considered.
  2. The experimental evaluation is relatively comprehensive. The authors test GLAFE on three EEG-fMRI datasets covering resting-state, auditory tasks, and neurofeedback, rather than relying on a single dataset. They also consider different ROI parcellations, multiple subject-disjoint splits, and cross-dataset transfer, including zero-shot evaluation. I particularly appreciate the use of multiple subject splits, since simultaneous EEG-fMRI datasets usually contain relatively few subjects, and the results can be sensitive to how the data is divided. They also compare with NeuroBOLT, UnEBOLT, REVE, and BIOT.
Critical weaknesses
  1. My first concern is whether the reported results really reflect ROI-specific brain activity. Since BOLD signals contain a shared global component, predicting this component well may improve temporal R and partly explain the FC results. I wonder how GLAFE and the baselines would perform if the global signal were regressed out from both predictions and ground truth. This would help clarify how much local information the model actually recovers. Since the central claim is that GLAFE recovers ROI-specific dynamics and connectivity, this evaluation is necessary to support the main contribution.
  2. I also have some doubts about the contribution of the proposed components. In Section 5.1, removing the global-local decomposition reduces temporal R by 0.056 and FC R by 0.042. From Table 6, this gives approximately 0.393 and 0.477, which are actually lower than the fine-tuned REVE baseline. This makes me wonder whether the improvement mainly comes from the global-local decomposition rather than the more complex cross-attention and BrainLM modules. I would be interested to see how a simpler REVE model with the same decomposition performs. Also, the improvements over the strongest baselines are sometimes small. As a result, two of the three proposed components are not shown to outperform a simpler baseline in the authors’ own ablation, which makes the methodological contribution unclear.
Questions for authors
  1. I am particularly interested in how well GLAFE recovers the local component of fMRI signals. Since the global signal is shared across ROIs, it may contribute substantially to both temporal prediction and FC reconstruction. Could the authors report temporal R and FC metrics after removing the global component from both predicted and ground-truth fMRI, and evaluate all baselines in the same way? I think this would help clarify whether the improvements mainly come from recovering the shared global activity or the ROI-specific fluctuations. It would also be useful to compare with simple baselines, such as predicting only the global signal or using the average FC matrix from the training set.
  2. I find the global-local decomposition interesting, but the current ablation also changes the supervision and prediction targets. Could the authors clarify how much of the improvement comes from the decomposition itself rather than the additional local supervision or loss functions? Also, would the same improvement hold across different subject splits and ROI parcellations?
  3. The autoregressive correction uses pretrained BrainLM and previous fMRI predictions, but removing this module only decreases temporal R by 0.025 and FC R by 0.009. Have the authors compared it with a simpler temporal model or the same module without BrainLM pretraining? I wonder how much of the improvement comes from BOLD temporal autocorrelation rather than the pretrained representations. Also, the fine-tuned REVE baseline already achieves FC R = 0.500 at 64 ROIs, compared with 0.521 for GLAFE. This makes me curious about how much of GLAFE's advantage comes from the pretrained REVE features rather than the proposed components.
  4. The FC losses are calculated using only 20 TRs per patch, even when there are 128 or 256 ROIs. Could the authors explain how reliable these short-window FC estimates are? Would using longer windows for the FC loss, while keeping the BrainLM patch structure unchanged, affect the results?
  5. I find the ROI-EEG attention maps in Figure 4 interesting, especially the HRF-like temporal pattern. However, I wonder whether the attention scores really reflect the importance of individual electrodes. Have the authors considered masking high-attention and low-attention electrodes at inference time to see how much the prediction performance changes?
5 comments

We thank the reviewer for their time, effort and insights. We address each of the weaknesses and questions as follows.

W1. My first concern is whether the reported results really reflect ROI-specific brain activity… I wonder how GLAFE and the baselines would perform if the global signal were regressed out from both predictions and ground truth… Since the central claim is that GLAFE recovers ROI-specific dynamics and connectivity, this evaluation is necessary to support the main contribution.

We evaluated all methods after regressing out the global signal from predicted and ground-truth ROI time series separately. GLAFE’s predictions of the local component consistently outperform all four baselines across all atlas sizes and evaluation metrics (temporal R, FC R, FC MSE). Its advantage therefore persists after removing the shared global component. Detailed global and local results appear in Tables 2, 10 and 11.

W2. I also have some doubts about the contribution of the proposed components… This makes me wonder whether the improvement mainly comes from the global-local decomposition rather than the more complex cross-attention and BrainLM modules. I would be interested to see how a simpler REVE model with the same decomposition performs…

We conducted a comprehensive ablation study, showing the contribution of each component. Following the reviewer’s suggestion, we evaluated a simpler model, retaining the global-local decomposition and REVE features, while replacing the cross-attention with a simple MLP and removing the autoregression correction module (Appendix B).

Compared to the ablated model, GLAFE achieves higher temporal R and FC MSE, alongside higher temporal R for both the local and global signal components separately. These results indicate that the cross-attention and autoregression mechanisms have a real contribution beyond the global-local decomposition.

Q1. I am particularly interested in how well GLAFE recovers the local component of fMRI signals… Could the authors report temporal R and FC metrics after removing the global component from both predicted and ground-truth fMRI, and evaluate all baselines in the same way… It would also be useful to compare with simple baselines, such as predicting only the global signal or using the average FC matrix from the training set.

As discussed in W1 above, GLAFE outperforms all baselines in the reconstruction of the local signals in both temporal and FC metrics. We further evaluated the two suggested heuristic baselines: a global signal-only baseline and a training-set population-average FC matrix (Table 9 Appendix D). Compared to the global signal alone, GLAFE’s full predictions achieve higher temporal R at every atlas size. This suggests that although the global component is the main driver of the dynamics, GLAFE can reconstruct local temporal information. Moreover, the global signal baseline naturally fails to reconstruct any connectivity pattern.

The population-average FC baseline achieves better FC metrics (FC R, FC MSE) than all prediction models, reflecting the substantial connectivity structure shared across individuals. However, predicting a static FC matrix is a considerably simpler task than reconstructing correlated time series that preserve such structure. Furthermore, To test whether GLAFE captures individual information beyond shared connectivity, we compared the FC R between predicted scans and their ground truth scans, to the FC R between the predicted scans and other subjects’ scans. Correlation with the corresponding scan was 10.87% higher on average, supporting the presence of individual information in GLAFE’s predictions beyond population average. See W4 in our comment to reviewer TVaj05.

Q2. I find the global-local decomposition interesting, but the current ablation also changes the supervision and prediction targets. Could the authors clarify how much of the improvement comes from the decomposition itself rather than the additional local supervision or loss functions? Also, would the same improvement hold across different subject splits and ROI parcellations?

We thank the reviewer for this clarification. We designed a new ablation, replacing the decomposed architecture with a single cross-attention pathway, while retaining the supervision unchanged. We applied the different components losses to the decomposed predicted full signal (Appendix B).

Compared to the new decomposition ablation, GLAFE achieves higher full-signal temporal R and local FC R at all three atlas sizes, while the ablation achieves lower full-signal FC MSE. These results support the contribution of explicit decomposition beyond retaining component supervision alone. Please see Tables 2, 10 and 11 for the full evaluation.

Q3. The autoregressive correction uses pretrained BrainLM and previous fMRI predictions, but removing this module only decreases temporal R by 0.025 and FC R by 0.009. Have the authors compared it with a simpler temporal model or the same module without BrainLM pretraining? I wonder how much of the improvement comes from BOLD temporal autocorrelation rather than the pretrained representations. Also, the fine-tuned REVE baseline already achieves FC R = 0.500 at 64 ROIs, compared with 0.521 for GLAFE. This makes me curious about how much of GLAFE's advantage comes from the pretrained REVE features rather than the proposed components.

We have expanded the autoregression-removal ablation to all three atlas sizes and subject splits (Appendix B). Full GLAFE achieves higher full-signal and local temporal R at every atlas size, averaged over splits. Moreover, neither causal moving-average nor Gaussian smoothing on the ablated model’s predictions recover the autoregression removal decrease in full and local temporal R (Tables 4-5).

Following the reviewer’s suggestion, we also compared pretrained and randomly initialized BrainLM. The following table shows the relative change when replacing pretrained weights with random initialization.

Atlas size (ROIs)Temporal R ↑FC R ↑FC MSE ↓
64+0.02%-0.14%+0.92%
128-0.09%-0.26%-0.91%
256-0.22%+0.51%+0.62%

The results show no consistent advantage to pretrained initialization, suggesting the correction architecture can achieve comparable performance under the same training without the pretrained checkpoints. Pretraining improved performance in our early experiments, but the correction mechanism has changed substantially since, and we thank the reviewer for prompting us to revisit this assumption. We are currently conducting further experiments to assess the finding’s robustness, and will update the explanatory claims accordingly in the final paper.

Regarding the contribution of the pretrained REVE features, GLAFE improves on fine-tuned REVE in 8/9 comparisons (Table 1), including 10.2–11.2% higher full signal temporal R and 20.9–33.9% higher local temporal R across all atlas sizes. Moreover, these temporal gains are statistically significant (p<0.05) across all atlas sizes (Appendix F). REVE provides strong EEG features and competitive FC reconstruction. However, GLAFE’s central objective is to reconstruct whole-brain fMRI time series while preserving correct connectivity. GLAFE’s temporal advantage shows that it is significantly better at this task, while still improving on FC reconstruction.

Q4. The FC losses are calculated using only 20 TRs per patch, even when there are 128 or 256 ROIs. Could the authors explain how reliable these short-window FC estimates are? Would using longer windows for the FC loss, while keeping the BrainLM patch structure unchanged, affect the results?

We thank the reviewer for highlighting an error in Appendix B.2, now corrected: Stage-2 FC losses use two concatenated patches (40 TRs), while Stage-1 uses 20 TRs. BrainLM’s patch structure remains 20 TRs throughout. We selected these horizons empirically to balance the trade-off between FC-estimate stability and sampling diversity, both subject to computational constraints.

Following the reviewer’s suggestion, we evaluated a version of GLAFE with 60-TR FC-loss horizons in both stages, retaining BrainLM’s structure. The three-split averages below show improved FC metrics but substantially lower temporal R at 128 and 256 ROIs, illustrating a trade-off between temporal and connectivity fidelity.

Atlas size (ROIs)ModelTemporal R ↑FC R ↑FC MSE ↓
64Full GLAFE0.352 ± 0.1420.521 ± 0.1160.111 ± 0.064
6460-TR horizon0.353 ± 0.1460.512 ± 0.1220.157 ± 0.068
128Full GLAFE0.341 ± 0.1360.520 ± 0.1170.141 ± 0.071
12860-TR horizon0.277 ± 0.1140.525 ± 0.1120.099 ± 0.056
256Full GLAFE0.303 ± 0.1300.562 ± 0.1020.169 ± 0.064
25660-TR horizon0.268 ± 0.1240.569 ± 0.1180.144 ± 0.048

Q5. I find the ROI-EEG attention maps in Figure 4 interesting, especially the HRF-like temporal pattern. However, I wonder whether the attention scores really reflect the importance of individual electrodes. Have the authors considered masking high-attention and low-attention electrodes at inference time to see how much the prediction performance changes?

We thank the reviewer for this suggestion. As a proof of concept, we conducted an exploratory inference-time electrode-selection experiment on the NeuroBOLT resting state dataset for one split at 64 ROIs. For each ROI, we retained only three of the 26 electrodes, selected using the electrodes attention scores from GLAFE’s spatio-temporal block. We compared the highest- and lowest-scoring eligble electrode triplets against five random triplets, applying spatial-separation and signal-correlation constraints to account for redudancy between electrodes.

For each inference run, predictions were evaluated for the ROI corresponding to the retained electrodes. We used the pretrained checkpoint from Table 1, without retraining. The following table shows the change in metrics beween the top and bottom triplets compared to the average result on the random triplets.

Selection vs. randomFull temporal R changeGlobal temporal R changeLocal temporal R change
Top-ranked−4.53 ± 3.76%−4.59 ± 4.02%+9.43 ± 18.97%
Bottom-ranked−5.08 ± 2.80%−4.42 ± 3.31%−1.07 ± 41.22%

Retaining top-ranked electrodes increased local temporal R by 9.43% relative to random selection, whereas bottom-ranked electrodes decreased local temporal R by 1.07%. Both selections performed worse than random on the global and full signals, possibly due to the need for large variation in the EEG signal to reconstruct the global signal.

Reviewer_CWXsOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces GLAFE, a model that predicts whole-brain ROI-level fMRI from simultaneous EEG. The main point is that BOLD signals share a dominant global component, so the model predicts the global signal and the local fluctuations of each ROI with separate pathways. The local pathway uses ROI embeddings that query EEG features through cross-attention, and an autoregressive step built on a fine-tuned fMRI model refines the prediction. It is evaluated on three datasets, three atlas sizes and three subject splits.

Critical strengths
  • The paper is well written.
  • The global/local decomposition is a simple and sensible idea.
  • The ROI-specific cross-attention to EEG features is interesting.
  • Better performance than the other models, evaluated over three splits.
Critical weaknesses
  • FC tends to over-represent the global signal (Murphy & Fox, NeuroImage 2017), so it is not clear how much of the performance comes from the global versus the local (residual) part. Only the full signal is reported.
  • Some baselines already do well enough. For example, fine-tuned REVE is close to GLAFE on FC R in Table 1 (0.554 vs. 0.562 at 256 ROIs).
  • The standard deviations in Table 2 are very large, so the gains over UneBOLT across datasets do not seem significant.
  • The autoregressive step (Sec. 3.3) is only described in words and would be clearer with equations.
Questions for authors
  • Please report global and local R separately, in Table 1 and in the ablations. Is the lost performance global or local?
  • Is removing the softmax necessary? What happens with it?
  • How does a Stage-1-only model compare to the full model? Is this the same as the no-autoregression ablation?
  • Does a larger atlas mean more ROIs, or smaller ROIs over the same brain? Temporal R goes down with more ROIs while FC R goes up.
2 comments

We thank the reviewer for their time, effort and insights. We address each of the weaknesses and questions as follows.

W1. FC tends to over-represent the global signal (Murphy & Fox, NeuroImage 2017), so it is not clear how much of the performance comes from the global versus the local (residual) part. Only the full signal is reported.

We evaluated the global and local components of the predicted signal separately for both FC and temporal reconstruction (Tables 2 and 10 in the revised manuscript). GLAFE’s predictions of both the local and global signal components consistently outperform all four baselines across all atlas sizes in both FC metrics (FC R, FC MSE), indicating that both components contribute to its performance.

W2. Some baselines already do well enough. For example, fine-tuned REVE is close to GLAFE on FC R in Table 1 (0.554 vs. 0.562 at 256 ROIs).

The goal of GLAFE is to accurately reconstruct whole brain fMRI dynamics, while faithfully preserving their connectivity patterns. Although REVE achieves a comparable FC R at 256 ROIs, GLAFE significantly outperforms it in temporal reconstruction, achieving 10.2%-11.2% improvement in temporal R, while also consistently improving FC R in all atlas sizes (3.5%-4.2%) and FC MSE in 64-128 ROIs (1.21%-7.31%). Inspecting the local signal reconstruction highlights this advantage (Table 2), where GLAFE outperforms all baselines in all metrics and atlas sizes, improving temporal R by 20.9–33.9% and FC MSE by 15.5–35.7%. These support GLAFE’s advantage as a predictor of whole-brain fMRI dynamics.

W3. The standard deviations in Table 2 are very large, so the gains over UneBOLT across datasets do not seem significant.

We evaluated the significance of GLAFE’s gains over UneBOLT in Table 2 (now Table 3) using paired permutation tests for each combination of dataset, training mode, atlas size, and evaluation metric, aggregating all three data splits (Appendix F). GLAFE outperformed UneBOLT in 49 of these 54 combinations. Among these comparisons, 25/49 yielded p < 0.05 and 32/49 had p < 0.10, providing statistical support for many of the reported gains.

Q1. Please report global and local R separately, in Table 1 and in the ablations. Is the lost performance global or local?

We added a separate evaluation of the local and global components, see Q1 in our response to reviewer TVaj, and Tables 2, 10 and 11 in the revised manuscript. We also added a separate global and local evaluation of the ablations in the revised ablation study (Appendix B). Removing the autoregressive correction, either alone or together with the cross-attention component, reduces local temporal R by 10.4–16.0% across atlas sizes, compared with only 0.5–4.3% reduction of global R, indicating the role of the autoregression in recovering local temporal dynamics. Removing the global-local decomposition results in smaller, mixed changes, as can be seen in the full ablation report.

Q2. Is removing the softmax necessary? What happens with it?

Omitting the softmax in the functional block is not explicitly necessary, but it is a beneficial design choice with an intuitive explanation. The functional block acts as a weighting mechanism for the ROI-conditioned EEG features, coming from the spatio-temporal block. It learns universal, input-independent weights for different combinations of ROIs and EEG feature types. Omitting softmax allows different features to receive positive or negative weights without the unnecessary constraint of a fixed total weight. To establish this experimentally, we evaluated a variant of GLAFE with softmax inside the functional block using the 64 ROIs atlas, resulting in worse local signal reconstruction, affecting full-signal FC prediction:

ModelFull temporal R ↑Full FC R ↑Full FC MSE ↓Global temporal R ↑Local temporal R ↑
Full GLAFE0.352 ± 0.1420.521 ± 0.1160.111 ± 0.0640.455 ± 0.1830.147 ± 0.057
With Softmax0.352 ± 0.1490.495 ± 0.1150.210 ± 0.1050.459 ± 0.1890.109 ± 0.058

Q3. How does a Stage-1-only model compare to the full model? Is this the same as the no-autoregression ablation?

No, a Stage-1 only model is not the same as the autoregression removal ablation. The goal of Stage-1 is mostly learning to predict the global signal, resulting in strong temporal fidelity but unfaithful FC reconstructions. In Stage-2, the GLAFE learns to predict the local component of the signal with minimal change to the global prediction, increasing FC reconstruction accuracy. This happens via two mechanisms in Stage-2: the autoregressive correction, and heavier FC-losses applied to the local signal predictions, while the global pathway remains frozen. Thus, a Stage-1 only model lacks more than simply the autoregression, and will exhibit weaker FC reconstruction abilities.

Q4. Does a larger atlas mean more ROIs, or smaller ROIs over the same brain? Temporal R goes down with more ROIs while FC R goes up.

In DiFuMo and many other functional atlases, increasing the dimensionality of the atlas results in both increased brain coverage and smaller ROIs on average (Dadi et al., 2020). The decrease in temporal R reflects the greater difficulty of predicting BOLD fluctuations of smaller regions. In contrast, FC R measures how well predictions reproduce the pattern of correlations between regions. Finer parcellations may reveal distinct connectivity patterns that are obscured in coarser resolutions by averaging over larger regions, potentially explaining the increase in FC R.

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

GLAFE is a whole-brain EEG-to-fMRI model that decomposes BOLD predictions into global, ROI-specific, and autoregressive components based on previous predictions. Its cross-attention mechanism uses learned ROI embeddings to predict 64 to 256 regions simultaneously from spatiotemporal EEG features. Across resting-state, auditory-task, and neurofeedback datasets, GLAFE achieves modest improvements over existing methods on temporal and functional connectivity metrics, while also showing promising zero-shot transfer and temporal attention patterns consistent with the canonical hemodynamic response.

Critical strengths
  1. The global-local design enables simultaneous whole-brain prediction across 64 to 256 ROIs.
  2. GLAFE performs consistently across datasets, atlas sizes, and training settings, with promising zero-shot transfer.
  3. The learned temporal and spatial attention patterns are consistent with known neurovascular responses and brain anatomy.
  4. Paper is very well-written and easy to follow.
Critical weaknesses
  1. Each ROI prediction includes a shared global component, and Figure 3 seems to show that most reconstructions follow brain-wide fluctuations common across ROIs. Could you report per-ROI temporal RR after regressing out the global signal from both predicted and measured series? A global-only baseline, with the global prediction broadcast to all ROIs, would also help establish how much of the performance comes from region-specific information.

  2. Could you report results without the autoregressive branch and compare against a non-autoregressive model with temporal smoothing? Also, does training use ground-truth past fMRI (teacher forcing) while inference relies on the model's own predictions? If so, could you show how performance changes over a 20-minute scan?

  3. GLAFE's temporal RR gains over the strongest baseline are relatively small (0.016 to 0.035), while its across-scan standard deviation ($\pm$0.13 to 0.14) is roughly twice that of NeuroBOLT and REVE. Could you report paired per-scan significance tests and explain this higher variance? Also, how was NeuroBOLT, originally a single-ROI model, adapted to 64 to 256 ROIs? Why are NeuroBOLT and REVE missing from Table 2, particularly given that REVE was the strongest FC competitor in Table 1?

  4. BIOT achieves an FC R=0.495R = 0.495 at 256 ROIs despite having the lowest temporal RR, suggesting that FC performance may largely reflect connectivity patterns shared across subjects. Could you compare all methods against a baseline that predicts the training-set group-average FC? It would also be useful to evaluate subject identifiability by checking whether each predicted FC matches its own subject's measured FC better than those of other test subjects.

I lean toward a weak accept, provided the authors can adequately address the concerns raised above.

Questions for authors

Please see the Weaknesses section.

4 comments

We thank the reviewer for their time, effort and insights. We address each of the weaknesses as follows.

W1. … Could you report per-ROI temporal R after regressing out the global signal from both predicted and measured series? A global-only baseline, with the global prediction broadcast to all ROIs, would also help establish how much of the performance comes from region-specific information.

We thank the reviewer for this suggestion. We evaluated per-ROI local temporal R after regressing out the global signal from both predicted and ground-truth fMRI time series (Table 2 in the revised submission). GLAFE outperforms all four baselines in predicting the local signals across all atlas sizes and evaluation metrics (Temporal R, FC R, and FC MSE), demonstrating its ability to recover region-specific dynamics beyond shared global fluctuations.

In addition, we evaluated the requested global-only baseline by broadcasting GLAFE’s predicted global signal to every ROI. Although the global component accounts for a substantial portion of the temporal correlation, GLAFE’s full predictions achieve higher temporal R at every atlas size (Table 9 in Appendix E), showing that reconstructed local dynamics contribute additional information beyond shared global fluctuations.

W2. Could you report results without the autoregressive branch and compare against a non-autoregressive model with temporal smoothing? Also, does training use ground-truth past fMRI (teacher forcing) while inference relies on the model's own predictions? If so, could you show how performance changes over a 20-minute scan?

Following the reviewer’s suggestion, we evaluated GLAFE without the autoregressive branch across all atlas sizes and data splits, including variants with temporal smoothing using moving average and Gaussian filters (Table 4, Appendix B). Compared to both variants, GLAFE achieves 1.5–9.4% higher temporal R and 1.7–23.6% lower FC MSE across all atlas sizes, demonstrating the contribution of the autoregressive branch beyond temporal smoothing.

Training did involve teacher forcing, but not exclusively. GLAFE was trained with a mixture of both ground-truth and generated past fMRI context windows, at a 70%-30% ratio, respectively (Appendix B). This mixed approach resulted in better performance than using either source alone. At inference, only model-generated past fMRI is used.

As requested, we added an evaluation of GLAFE’s within-scan performance, plotting temporal R as a function of time (Figure 6, Appendix E). All methods show an overall increase in temporal R as the scan progresses, where GLAFE generally follows the same trajectory, with a slightly more pronounced improvement during the later portions of the scan.

W3. GLAFE's temporal R gains over the strongest baseline are relatively small (0.016 to 0.035), while its across-scan standard deviation (0.13 to 0.14) is roughly twice that of NeuroBOLT and REVE. Could you report paired per-scan significance tests and explain this higher variance? Also, how was NeuroBOLT, originally a single-ROI model, adapted to 64 to 256 ROIs? Why are NeuroBOLT and REVE missing from Table 2, particularly given that REVE was the strongest FC competitor in Table 1?

We thank the reviewer for highlighting the standard deviation (SD) issue. We found an error in the reported SD for BIOT, NeuroBOLT, and REVE: we reported within-scan SD across ROIs, instead of the across-scans SD, averaged over ROIs. We and have corrected the reported SDs in the revised submission. After correction, GLAFE’s standard deviations are comparable to those of the evaluated baseline methods (Table 1 in the revised manuscript).

We assessed the significance of GLAFE’s temporal R gains over the strongest baseline at each atlas size using paired permutation tests, combining evidence across all three data splits (Appendix F). The resulting p-values are 0.075, 0.0044, and 0.176 for 64, 128, and 256 ROIs, respectively.

Regarding NeuroBOLT’s adaptation, although its paper (Li et al., 2024) introduces only a single-ROI version of the model, its official code includes a multi-ROI variant as well. We followed this implementation, setting the final prediction head’s output dimension to the number of ROIs. We initially focused on UnEBOLT in Table 2 because it explicitly evaluates whole-brain EEG-to-fMRI prediction under all three settings: training from scratch, fine-tuning, and zero-shot. Following the reviewer’s suggestion, we have initiated the corresponding REVE and NeuroBOLT experiments, which are currently ongoing. We we will update the revised manuscript with the results as they become available.

W4. BIOT achieves an FC R=0.495 at 256 ROIs despite having the lowest temporal R, suggesting that FC performance may largely reflect connectivity patterns shared across subjects. Could you compare all methods against a baseline that predicts the training-set group-average FC? It would also be useful to evaluate subject identifiability by checking whether each predicted FC matches its own subject's measured FC better than those of other test subjects.

Following this suggestion, we evaluated the training-set population average FC as an additional baseline. It achieves the highest FC R and lowest FC MSE across all methods (Table 9 in Appendix E), but addresses a much simpler task: predicting connectivity alone. GLAFE must reconstruct ROI time series that capture both temporal dynamics and connectivity. Following the reviewer’s suggestion, we inspected whether the predicted FC matrices correlate more with their own ground truth counterparts, compared to other subjects’ FC matrices, as can be seen in the table below.

Atlas size (ROIs)FC R: other subjectsFC R: same scanRelative increase
640.445 ± 0.0890.521 ± 0.116+17.11%
1280.468 ± 0.0680.520 ± 0.117+10.98%
2560.538 ± 0.0350.562 ± 0.102+4.53%

Averaged over all atlas sizes and splits, same-scan FC correlations are 10.87% higher than correlations with other subjects’ FCs, suggesting GLAFE learns individual connectivity patterns beyond population averages.

Reviewer_TVajOpenReview ↗

Reply to Authors Rebuttal

I thank the authors for a thorough rebuttal, I maintain my score of weak acceptance. A few remaining points:

(1) The group-average FC baseline outperforming all models on FC metrics is an important result and should appear in the main text, not only the appendix. For identifiability, please add a significance test, the same analysis for the baselines, and either top-1 identification accuracy or a correlation between predicted and measured deviations from the group-average FC. (2) Please include the REVE and NeuroBOLT results for Table 2 once available. (3) Given the BrainLM-initialization and electrode-masking results reported to other reviewers, please revise claims about the role of pretrained BrainLM and the interpretability of the attention maps.

We greatly thank the reviewer for reading and assessing our response. We have uploaded a revised manuscript with the additional REVE/NeuroBOLT results for Table 2. We are currently working on a revised manuscript that will include all new evaluations in the main text. Please see our new general comment for details.

Comments on the paper

2 comments

We thank all reviewers for their greatly appreciated feedback. We have uploaded a revised version of the manuscript.

The main additions include:

  • Separate evaluations of global and local signal reconstruction for all methods

  • An expanded ablation study across atlas sizes and subject splits, covering the full signal and its global and local components

  • Two additional heuristic baselines: global-signal prediction and training-set average FC

  • Significance tests for all comparisons in Tables 1 and 2, across all datasets, atlas sizes and metrics

We have uploaded a revised manuscript with the completed evaluations of REVE and NeuroBOLT as additional baselines in the generalization across datasets task and have added them to the main text of the revised submission (Table 3).

In the new evaluations, GLAFE outperforms all baselines in 43 out of 54 comparisons across datasets, training modes, evaluation metrics, and atlas sizes. Specifically, in temporal R and FC R, GLAFE outperforms all baselines in 17/18 and 15/8 comparisons, respectively.

We are currently working on a final version of the paper which will include a revision to the claims regarding BrainLM pretraining, a more nuanced discussion of the potential practical uses of the electrodes attention maps, and the adding the new controls (population average FC and global signal only) to the main text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.