- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This work proposes BOLDFlow, a novel conditional flow-matching framework that generates ROI-level BOLD time series from EEG and computes functional connectivity from the generated signals. The proposed method combines a pretrained REVE encoder with multiscale spectral features and a learned EEG-conditioned Gaussian source. They also asked the. model to preserve cross-region. covariance that may be lost in previous regression-based supervision. Experiments across two concurrent EEG–fMRI datasets show significant improvements in both temporal correlation and FC recovery, with comprehensive analyses covering different aspects, such as the community structure of the recovered FC, scalability to finer brain parcellations, and uncertainty estimation.
Critical strengths
- The paper presents extensive experiments covering EEG context lengths under both seq2one and seq2seq settings, electrode configurations, and architectural and training variants. Subject-disjoint cross-validation with multiple random seeds strengthens confidence in the reported results.
- The proposed method consistently outperforms the evaluated baselines on both datasets, with particularly substantial improvements in FC recovery, while also improving the temporal correlation between pred and g.t..
- The results and discussion of the trade-off between temporal reconstruction accuracy and FC recovery is insightful.
Critical weaknesses
-
The writing and presentation could be improved. A clearer explanation of the prediction and aggregation pipeline, particularly the distinction between averaging overlapping block predictions and averaging independently sampled trajectories, would make the method easier to follow. The figures would also benefit from improved legibility and clearer illustration of the complete pipeline. The work provided extensive experimental evaluation, which is great, but the presentation sometimes feels fragmented. Better organization and a clearer synthesis/summary of the key findings may help understand how the individual experiments support the paper’s main claims.
-
Although the method is evaluated on two datasets, training and testing are performed separately within each dataset. These experiments support the effectiveness of the method across datasets when retrained, but do not establish the transferability of a trained model across datasets or acquisition conditions, which is important for this task and broader application of the method.
-
The methodological novelty appears a little bit limited. The framework largely combines existing EEG encoders, conditional flow matching, and β-NLL. Although the empirical results are strong, the contribution appears primarily empirical rather than a substantial methodological advance.
Questions for authors
-
Table 5 results is intereting. If I understand correctly, this is the comparison between seq2one and seq2seq with T=4 averaging. Have you compared, using the same trained seq2seq model, how do temporal correlation and FC correlation change when each TR is reconstructed using a single prediction from a fixed position within the output block (i.e., seq2seq but T=1, maybe just choosing the last prediction point), rather than by averaging overlapping predictions?
-
How was the Tout = 4 been selected, or have you compared with other numbers e.g., Tout = 2, 6 etc?
-
FC recovery is the primary evaluation target, yet the training objective supervises BOLD sequence generation without an explicit FC-based loss. Have you considered or evaluated adding an auxiliary FC loss if it's feasible?
-
Beyond the NeuroBOLT-encoder hybrid in Table 1, have you compared REVE with other EEG foundation-model backbones under otherwise matched settings?
4 comments
Comment 2/2: Questions
[Q1] Single-position readout (seq2seq without overlap averaging).
Your understanding of Table 5 is exact. To investigate your point, we took the same trained seq2seq model and reconstructed each TR from a single fixed position of the output block (either the last or the first prediction point) instead of averaging:
| Readout | T. Corr | FC Corr |
|---|---|---|
| Seq2one, 32 s | 0.317 | 0.564 |
| Seq2seq, single position: last TR | 0.313 | 0.575 |
| Seq2seq, single position: first TR | 0.315 | 0.569 |
| Seq2seq, overlap average (reported) | 0.326 | 0.584 |
Without overlap averaging, single-position seq2seq performs on par with seq2one, while averaging overlapping blocks yields a consistent gain in both T. Corr and FC Corr. This is in line with the 26–32 s context plateau in Table 5 (seq2one T. Corr: –), since both output positions still see of preceding EEG.
[Q2] Selection of .
We fixed without a grid search over other block lengths, since seq2one () already achieves FC Corr and tuning is orthogonal to our core generative contribution. Practically, is coupled to the input window length (App. C): larger blocks leave less preceding EEG for the earliest TR unless the input window is also extended, and provides overlap averaging while keeping all outputs within the hemodynamic regime at (Table 5).
[Q3] Adding an auxiliary FC loss. We considered adding an explicit FC loss early on, but decided against it because it does not fit cleanly into per-window flow-matching training. Each training step only predicts the velocity of one 4-TR block at a random flow time , without integrating the ODE to generate BOLD outputs. Since FC is a correlation over a full scan (Eq. 1), adding an FC loss would require batching long consecutive segments from the same scan and unrolling ODE integration inside the training loop. Moreover, learning the conditional distribution already targets the second moments that determine FC (Corollary 1), lifting FC Corr from to without an explicit FC term (Table 6).
[Q4] Comparing REVE with other EEG foundation-model backbones. Our framework is encoder-agnostic, and Table 1 already isolates the backbone effect by pairing our generative decoder with two distinct EEG foundation models, LaBraM (via NeuroBOLT's encoder) and REVE: the decoder alone drives the bulk of the FC gain ( on NeuroBOLT; on OpenNeuroSleep), while upgrading the backbone to REVE adds a consistent / . We selected REVE as our default rather than sweeping additional backbones because it reports state-of-the-art results on ten downstream EEG tasks, with an average balanced-accuracy gain of over the previous best foundation model, CBraMod (El Ouahidi et al., 2025), and its 4D coordinate encoding natively handles arbitrary montages (26, 30, 19, and 6 channels; App. C) without architectural changes.
Thank you again for your time and constructive feedback. We hope our clarifications and new experiments address all of your questions, and we would be grateful if you considered raising your score.
Thank you again for your constructive feedback. We have clarified the prediction and averaging pipeline and added the single-position readout and cross-dataset experiments discussed in our response. We hope these additions help, and we would welcome any further questions or remaining concerns.
Thank you for the detailed response and revisions. Most of my concerns have been addressed, so I am happy to raise my score.
Comment 1/2: Weaknesses
Thank you for your thorough review and for recognizing our novel conditional flow-matching framework, rigorous multi-seed subject-disjoint evaluation, consistent gains in FC and temporal correlation, and insight into the accuracy–FC trade-off.
Revisions in the PDF are highlighted in blue. Below we address your concerns (Part 1/2: Weaknesses; Part 2/2: Questions):
[W1] Clarification on the pipeline, figures, and organization. We describe the pipeline and both types of averaging in the Method section (with pseudocode in Algorithm 1, App. B), and we made this comparison more direct in the revision. Both
seq2oneandseq2seqtake the same 32 s EEG window as input; the only difference is thatseq2oneoutputs one BOLD vector (1 TR), whileseq2seq(our default) outputs a matrix of 4 TRs. Inseq2seq, averaging overlapping blocks takes the mean of overlapping 4-TR predictions obtained by shifting the EEG window by 1 TR to form a single trajectory (M=1, Eq. 8; used for all Table 1 results), whereas averaging independent trajectories runs this generation M times with fresh noise draws and averages the M trajectories (used only for the ensemble ablation in Table 14 and UQ at M=50).Regarding figure legibility, we redrew all figures with larger fonts and restructured Figure 2 into a clear end-to-end pipeline diagram.
Finally, we appreciate your note on the breadth of our experiments. To make the Experiments section easier to follow, we added a short opening paragraph linking its three parts (connectivity recovery, decoder/readout analysis, and uncertainty estimation) directly to our main contributions and findings.
[W2] Analysis on the transferability across datasets and acquisition conditions. We do not claim cross-dataset generalization of a trained model (noted as a limitation in App. I). Instead, as you note, we test the transferability of the method: freezing all architecture and hyperparameters on NeuroBOLT and training unchanged on OpenNeuroSleep, where BOLDFlow still outperforms all baselines (0.340→0.528 FC Corr, Table 1). Clean zero-shot evaluation on public paired data is difficult because NeuroBOLT (N=22) and OpenNeuroSleep (N=33) differ simultaneously in scanner, EPI sequence, montage, brain state (rest vs. sleep), and age.
Following your suggestion, we trained BOLDFlow on OpenNeuroSleep alone (22 shared channels) and tested it zero-shot on NeuroBOLT with no fine-tuning: it reaches 0.528±0.108 FC Corr, beating in-domain NeuroBOLT (0.428) and UnEBOLT (0.418), whereas T. Corr drops to 0.130. These preliminary results suggest that BOLDFlow can generalize across datasets, with a clear takeaway: connectivity structure transfers zero-shot across scanners and brain states, whereas pointwise timing shifts without light target-domain adaptation (App. F).
[W3] Clarification on methodological novelty and combining existing components. We are glad you found our empirical results strong, and while individual building blocks (REVE, CFM, β-NLL) exist, simply combining off-the-shelf parts does not recover FC. To our knowledge, no prior EEG-to-fMRI work has diagnosed why point predictors fail on FC or formulated connectivity recovery as conditional distribution matching; our core contribution is identifying this fundamental limitation and redesigning the generative pipeline around three critical bottlenecks: