acceptodds
Withdrawn from ICLR 2027

Learning Lead Field Corrections for EEG–fMRI Functional Connectivity Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Joint analysis of EEG and BOLD-fMRI signals is a route to studying how brain structure relates to brain function. This paper presents a model that learns a low-rank correction to the T1-derived lead field per subject. The correction is fitted by matching the EEG-side parcel covariance and the fMRI-derived parcel covariance, each estimated within-subject from the corresponding modality's own resting data. We call this construction Lead-Field Low-Rank Correction (LF-LoRA). The covariance matching has the form , where represents the EEG/BOLD-parcel signal covariance and is a linear EEG-to-parcel operator built from the corrected lead field . The simple model structure is consistent with the linear forward-modeling assumption from which the lead field is defined, and its solution may help explain subject-specific structure in the EEG forward model (Hämäläinen et al., 1993). The linear model admits a decoupled temporal operator which aligns and at lag . The temporal operator provides a means to investigate the dynamics of neural activity from EEG. We discuss the solution space of LF-LoRA. We evaluate the construction on the LEMON cohort (Babayan et al., 2019). Experiments validate the subject specificity of the learned model and the utility of reading on cortex.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, fMRI, lead field, forward model, low-rank correction, functional connectivity

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_LCSkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Lead-Field Low-Rank Correction (LF-LoRA), a subject-specific framework for aligning EEG-derived functional connectivity (FC) with fMRI-derived FC. The method introduces a learnable low-rank correction to the anatomically derived EEG lead-field matrix and optimizes it by minimizing the discrepancy between EEG-reconstructed parcel-level covariance and fMRI-derived functional connectivity. The authors further extend the framework with a linear temporal operator to incorporate lagged covariance information.

Experiments are conducted on the LEMON dataset with 203 subjects. The proposed method achieves an average cross-state EEG–fMRI FC correlation of approximately 0.63, outperforming conventional source reconstruction and EEG connectivity estimation methods. Additionally, the method improves subject fingerprinting performance, increasing verification AUC from 0.912 to 0.950. The paper also presents temporal alignment experiments and visualizations of the learned corrections to investigate their spatial and temporal characteristics.

Critical strengths
  1. The paper addresses an important challenge in multimodal neuroimaging by leveraging fMRI-derived functional connectivity to calibrate EEG lead-field models. This provides a promising direction for improving EEG-based brain connectivity analysis.

  2. Experiments on 203 subjects from the LEMON dataset demonstrate substantial improvements in EEG–fMRI functional connectivity alignment over conventional baselines. The method also improves subject fingerprinting performance, increasing verification AUC from 0.912 to 0.950.

Critical weaknesses

critical weakness:

  1. All real-data experiments are conducted on a single dataset (LEMON), with evaluations restricted to eyes-open and eyes-closed resting-state conditions. Although the study includes 203 subjects, the lack of validation across different datasets, cognitive tasks, and acquisition settings limits the evidence for the generalizability of the proposed method. Furthermore, despite introducing a temporal dynamics modeling component, the paper does not evaluate it on simultaneously acquired EEG–fMRI data or provide independent validation of lagged covariance prediction. These limitations weaken the experimental support for the method's broader applicability and temporal modeling capabilities.

  2. The proposed method is primarily compared against conventional approaches such as MNE and eLORETA, which do not receive the same fMRI supervision as LF-LoRA. Consequently, it remains unclear whether the reported improvements arise from the proposed low-rank lead-field correction or simply from the additional cross-modal supervision. Moreover, the paper lacks systematic comparisons with alternative learnable linear mappings under matched supervision, as well as comprehensive ablation studies examining the contributions of anatomical initialization, low-rank constraints, and the temporal operator. The existing hyperparameter ablations are conducted on only 30 subjects with a single random seed, without sufficient evaluation of experimental stability. These limitations make it difficult to establish the necessity and distinct advantages of the proposed framework.

Questions for authors
  1. Improved FC alignment does not necessarily indicate a more accurate physical lead-field model. How do the authors demonstrate that the learned corrections are physically meaningful rather than simply fitting fMRI connectivity patterns?

  2. Appendix D indicates that cross-state FC alignment performance is used for checkpoint selection. How do the authors prevent potential selection bias in the final evaluation?

  3. EEG and fMRI in LEMON were not acquired simultaneously and exhibit different temporal characteristics. How do the authors justify the lagged covariance alignment and validate the physiological significance of the temporal operator?

  4. What are the advantages of applying low-rank corrections to the lead-field matrix rather than directly learning an EEG-to-parcel mapping? Could the authors provide comparisons under equivalent supervision?

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

The authors propose a low-rank scheme for optimizing anatomically derived EEG lead fields by aligning them with fMRI functional connectivity. This is performed by adding a low-rank correction to the lead field matrix and optimizing the discreepancy beween EEG-derived connecitvity and fMRI connectivity. The authors perform several experiments to demonstrate the improvement to EEG-fMRI alignment, subject specific fingerprinting, and lagged covariance matching between EEG and fMRI. These experimemtns are demonstrated on N=203 subjects from the LEMON dataset. There are several interesting ideas in the paper, however, there were a few questions about the validity of the claim, and the exact applicability of the approach.

Critical strengths

The topic is important. The ideas proposed in the paper are simple and elegant. This is one of the strengths. To my knowledge, this is one of the first works that have proposed the low-rank correction of EEG-derived connectivity under resting state fMRI supervision.

The idea of constraining EEG with fMRI itself is not new, and there have been several attempts to improve the EEG source reconstruciton with fMRI connectivity (the authors have cited some of these works.). However, the additive low-rank approach to learn lead field corrections (A_0 + UV^T) is a novel idea in the EEG-fMRI field for imposing structure on the optimization problem.

The two-stage optimization strategy (Appendix G) is a clever way of solving the problem. Although individual elements of this optimization strategy (GL(r) group action, searching for an element in the ortbit of the Stiefel manifold) have been well studied before, the idea of combining these ideas, and applying it to the EEG-fMRI problem is a good contribution.

The experimental results demonstrate the effectiveness of the method and also demonstrate cross-transfer of learned correction between different EEG experimental conditions. The fingerprinting results are strong (Top-1 accuracy reaching 100%). Figure 2 is one of the main contributions of this paper.

Critical weaknesses

While there are a few novel aspects to this method as regards the field of EEG-fMRI connectivity, there are several weaknesses related to how the method is evaluated and how exactly can the method be applied to common applications.

Because the authors set to align the EEG and fMRI connectivity, by design they are going to show an incrase in the mean Pearson r. For e.g. Table 1 shows a drastic magnitudes of order improvement over methods such as MNE and eLORETA, going from r=0.05 to r=0.63. This is fine and is a validation that the alignment itself worked. However, the paper doesn't do a good job of explaining how and where this alignment will be useful. Figure 2 is the only experiment that shows that the alignment of EEG and fMRI translates into a meaningful improvement of subject specificity (via fingerprinting). They show the AUC improving from 0.912 for MNE/eLORETA (which have not been trained on fMRI) to 0.95 for LF-LoRA. However at baseline (\alpha=0), MNE has achieved strong performance at 0.91 already.Thus beyond fingerprinting, the authors do not demosntrate any performance gains in other downstream tasks such as prediction, clinical classification etc. Thereby, the broader applicability of the method is not demonstrated.

The authors don't compare their method with other fMRI-informed EEG methods. Thus while I believe the idea of low-rank correction to EEG using fMRI is still novel in the field, the authors should at least discuss or compare other low-dimensional forward modeling methods. For e.g., the authors should cite and discuss these works Henson, Richard N., et al. "A Parametric Empirical Bayesian framework for fMRI‐constrained MEG/EEG source reconstruction." Human brain mapping 31.10 (2010): 1512-1531. Lei, Xu, et al. "fMRI functional networks for EEG source imaging." Human brain mapping 32.7 (2011): 1141-1160. Hansen et al. (2016). Data-Driven Forward Model Inference for EEG Brain Imaging. NeuroImage. Abreu, Rodolfo, et al. "Optimizing EEG source reconstruction with concurrent fMRI-derived spatial priors." Brain topography 35.3 (2022): 282-301. The authors do cite other papers from the groups of Abreeu et al. and Lei et al., however, above are more directly relevant to this paper. I would note that the above works are not exactly trying to solve what the authors propose here, but the ideas of EEG-fMRI integration and forward modeling are relevant to this topic. The authors compare their method to ConvDip, however it is not trained on fMRI.

There is now a well-established issue of source leakage. See for e.g. Colclough, Giles L., et al. "A symmetric multivariate leakage correction for MEG connectomes." Neuroimage 117 (2015): 439-448. Nagy, Péter, et al. "The effects of spatial leakage correction on the reliability of EEG‐based functional connectivity networks." Human Brain Mapping 45.8 (2024): e26747.

6. Potential effects of source leakage on functional connectivity and fingerprinting

This was not mentioned in the paper. It woould be interesting to see if the proposed method does better than the methods that specifically accont for this leakage. I.e. does fMRI have enough information that the proposed low-rank correction improves upon this problem. This also has implications for subject identifiability (that the authors tackle in this paper) and fingerprinting (which the authors also show in this paper).

The authors show results on temporal extension, which are quite interesting. However, as above there are no comparisons against other methods that estimate these lagged covariances matrices.

It was unclear what the authors mean by "sessions" in the paper. Are these repeated measurements on different days (i.e. longitudinal scans) or are these performed in the same single setting of the acquisition? The authors show results of stability. But another intesteing experiment would be to scan fMRI once, perform the connectivity alignment, and show stability across different sessions (different days) to see whether one has retained enough information in the model.

Questions for authors

The main question here is whether the improved alignment leads to better performance improvements in other unrelated tasks that rely on EEG signal acquisition, processing, and analysis.

As stated above, how would the methods compare to those that use low-dimensional representations from fMRI to constrain EEG? For e.g. how is the low-rank correction to the lead field advantageous compared to the these other low-dimensional EEG-forward modeling approaches (Hansen et al. 2016)?

The EEG and BOLD signals occur at different time scales. While I understnd the need for aligning their respective connectivities, however, it would be important to discuss, what information is being retain and what is being discarded in this approach.

How is the recording session defined? Are eyes-open and eyes-closed separate sessions? Or different conditions during the same session? ARe these over multiple visits or a same visit?

Reviewer_zGTwOpenReview ↗
Rating
1
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LF-LoRA, a subject-specific low-rank correction to a T1-derived EEG lead field, learned by matching EEG-derived parcel covariance/functional connectivity to the same subject’s fMRI-derived connectivity. It also introduces a lagged temporal operator to align delayed EEG/fMRI covariance structure and evaluates the method on 203 healthy participants from LEMON.

Critical strengths
  1. Interesting multimodal formulation: The paper gives a mathematically clean way to incorporate subject-specific fMRI information into EEG source modeling through a low-rank correction rather than replacing the physical lead-field model entirely.
  2. Large public cohort and interpretability: Evaluation on 203 LEMON subjects is reasonably substantial, and the paper attempts to interpret both the learned spatial correction and lagged temporal operator through cortical visualizations.
Critical weaknesses
  1. Evaluation is potentially circular because the subject’s own fMRI FC is used to train the correction and then to judge alignment. The model is explicitly optimized so that BΣyB⊤B\Sigma_yB^\top approaches that same subject’s fMRI FC, and the cross-state experiment changes EEG state but retains the same subject-specific fMRI target. This demonstrates consistency across EEG states, but not independent prediction of unseen fMRI connectivity
  2. The baselines are not sufficiently strong for the main claim. The quantitative comparison is mainly against MNE, eLORETA, and envelope-correlation FC. All the methods are quite old. ConvDip is discussed only through a single-subject “smoke test,” rather than being evaluated systematically across the full cohort. A stronger comparison with modern learned EEG source-localization or EEG–fMRI alignment methods would be needed.
  3. Robustness/generalization evidence is limited. All experiments use one healthy dataset, one atlas, and effectively one main training configuration. The appendix explicitly states that multi-seed experiments and an alternative-atlas sanity check were not performed. The paper itself also acknowledges that clinical utility and patient-population validation remain untested.
Questions for authors
  1. How can the authors rule out circular evaluation when the same subject-specific fMRI FC is used both to train the correction and to assess EEG–fMRI alignment?

  2. Why are stronger modern learned EEG source-localization or EEG–fMRI alignment baselines not evaluated systematically across the full cohort instead of relying mainly on older methods and a single-subject ConvDip smoke test?

  3. How can the authors support robustness and generalization claims when the evaluation is limited to one healthy dataset, one atlas, a single main configuration, and no multi-seed, alternative-atlas, or patient-population validation?

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

This paper proposes refining the forward model of EEG source localization (lead field) by using a Low-Rank correction that optimizes for covariance matching with fMRI functional Connectivity, taking potential temporal lag in the signal into account. The method is trained and validated at the subject level using the LEMON-Cohort.

Critical strengths

Overall, the modeling and linear algebra framing is strong and well-defined, backed by a companion paper (Submission #54325)

Critical weaknesses

1.No evaluation holds out the fMRI target: LEMON provides a single fMRI run per subject, so B_s Sigma_y Bs^T is scored against the same Sigma ref it was trained on (4.2). Similarly, the Multi-lag (4.3) results are reported on the fitted data, as for the fingerprinting (4.4) results. The gains over MNE/eLORETA and envelope baselines thus largely reflect supervision on the target, rather than a better source map.

  1. The claim that the prediction depends on the subject’s EEG is never tested. The per-subject correction is large relative to the unique FC entries (V has 82k params; Sigma_y has only 11k). Using Sigma_y as an explicit factor argument does not preclude overfitting to the template solution, since Sigma_y shares the same montage and is trace-normalized across subjects. The experiments used against ConvDip (e.g., time-shuffling, noise replacement) have not been tested on Lf-LORA.
Questions for authors

Clarifications 1a. Test on held-out data: Split the fMRI run in half: LF-LoRa fitted on the FC of the first half of fMRI functional connectivity (FC); how does it score against the FC of the other half? Expected pass would be r to stay well above the unsupervised baseline and close to the noise ceiling of the split fMRI correlation. 1b. For fingerprinting, the subject's identity can leak through operator B. The same B_s should be used on both sides, and independently of it. Additionally, the classification is highly imbalanced, and alternative metrics to ROC-AUC are desirable (e.g., balanced accuracy, F-score). Additionally, B(alpha) expression is never explicitly stated.

2a. Perform the same experiment used against ConvDip: At test time, with Sigma_y from time-shuffled EEG, another subject's EEG, a cohort mean, and noise, the scores should drop significantly. 2b. EEG-free baseline: report “r” for the Cohort-mean fMRI FC (test subject held out) and against another subject's FC; LF-LoRa should exceed both controls.

  1. B(alpha): what is its explicit formulation?

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

Related papers

Discussion (0)

Sign in to comment.