Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

EEG emotion recognition faces significant hurdles due to noise interference, signal nonstationarity, and the inherent complexity of brain activity which make accurately emotion classification. In this study, we present the Fourier Adjacency Transformer, a novel framework that seamlessly integrates Fourier-based periodic analysis with graph-driven structural modeling. Our method first leverages novel Fourier-inspired modules to extract periodic features from embedded EEG signals, effectively decoupling them from aperiodic components. Subsequently, we employ an adjacency attention scheme to reinforce universal inter-channel correlation patterns, coupling these patterns with their sample-based counterparts. Empirical evaluations on SEED and DEAP datasets demonstrate that our method surpasses existing state-of-the-art techniques, achieving an improvement of approximately 6.5% in recognition accuracy. By unifying periodicity and structural insights, this framework offers a promising direction for future research in EEG emotion analysis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0161_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/YanhaoHuang23/FAT

Link to the Dataset(s)

SEED dataset: https://bcmi.sjtu.edu.cn/~seed/ SEED-IV dataset: https://bcmi.sjtu.edu.cn/~seed/seed-iv.html SEED-V dataset: https://bcmi.sjtu.edu.cn/~seed/seed-v.html SEED-VI dataset: https://bcmi.sjtu.edu.cn/~seed/seed-vii.html DEAP dataset: https://www.eecs.qmul.ac.uk/mmv/datasets/deap/

BibTex

@InProceedings{WanJin_Anovel_MICCAI2025,
        author = { Wang, Jinfeng and Huang, Yanhao and Song, Sifan and Wang, Boqian and Su, Jionglong and Ding, Jiaman},
        title = { { A novel Fourier Adjacency Transformer for advanced EEG emotion recognition } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {12 -- 22}
}

Reviews

Review #1

Please describe the contribution of the paper

This paper presents Fourier Adjacency Transformer (FAT), a novel transformer-based architecture for EEG emotion recognition. The key contributions are Fourier Analytic and Fourier Adjacent Transformer.

The results on public benchmark for EEG emotion recognition shows improved performance
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

Using Fouried Attention Layer for periodicity modeling and extensive benchmarking across five EEG datasets
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

No clinical or biological/translation analysis, just one visualization figure. Lack of statistical significane. Discussion is missing how Fourier Adjacency Transformer and Fourier Analytic Layer extracts the required informations No mention of open-source code for reproducibility
Please rate the clarity and organization of this paper

Satisfactory
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

No mention of open-source code for reproducibility. It will be difficult to reproduce using only the information from the paper.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(4) Weak Accept — could be accepted, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

The paper proposes an interesting idea with some good results, but in-depth analysis, discussion, and statistical significance are missing.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #2

Please describe the contribution of the paper

The paper proposes a new transformer layer suitable for processing EEG in the standard transformer architecture. The layer, called FAL, is particularly adapted to handle EEG data for emotion recognition, which the authors evaluate on a set of benchmark datasets.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

The method is evaluated on 6 different EEG benchmark datasets, includes a performance comparison for different frequency bands for three of these benchmarks, and includes an ablation study for adding or dropping the additive adjacency matrix and varying the output dimensionality ratio.
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

– The method is quite incremental compared to Fourier Analysis Networks, effectively changing only a minor detail of the algorithm. However, it seems that FAN has not been extensively benchmarked in the context of EEG analysis (?). It would be helpful if the authors better explain their contribution to the existing literature in this regard. – Assuming that in Table 1 and 2, the authors denote “Mean / Std”, the variation in the results is quite considerably. While the method improves across all prior methods, it is unclear if this improvement is statistically significant. It would be helpful if the authors could comment on this point, and potentially run a significance test. IF the variance is this high because the splits of the data are so different, it would be helpful to average out the effect of data splitting, and report a mean and standard deviation over re-runs of the model as well, to judge the difference between dataset differences and model differences. – Following up on the previous two points, given that FAL is a special case of FAN, it is unclear to me in table 1 whether the change gives actually better results. Could it be that with better parameter tuning or a larger dataset, or some other tweak of the model, the more flexible architecture performs just as well? – The evaluation and discussion of Fig 3 (Sec 3.5) is quite weak – what is the evaluation/scientific interpretation here? Such an analysis could be interesting, but it should be applied to a dataset where the basic structure of the expected pattern is known in advance, as a means to validate the approach. What do other approaches predict in this instance? – Overall, the presentation of the results is quite weak. The tables are not neatly typeset (a lot of clutter and lines), a lot of numbers are presented, the table captions are too short to parse the tables without looking into the text what eg the meaning of the dash “/” should be, how many repeats have been run, etc.; i.e., crucial info to actually interpret the significance of the results. – The benchmarking setting seems quite narrow. What about other EEG tasks? Or are the current datasets already so large that it is likely that the performance translates to other settings? Can the authors envision EEG tasks in which the architecture will be a suboptimal choice? Are the benchmarks used in the paper similar to the benchmark in the regarded baselines?
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

– Will a reference implementation of the method become available upon publication?

p.2, “recently, transformed-based models have demonstrated … […] However, their data inefficiency and lack of inductive bias hinder stability and generalization in EEG applications –> do you have evidence for this causal statement / or a suitable citation?

Eq (1), “B_n”: Bias term, would lower case. It would also be good to mention that the cos/sin are applied elementwise

Which results in Table 1/2 are reproduced by your experimental suite? Do the numbers presented for FAN match the ones reported in the original paper?

How were the hyperparameters for the approach selected?

It looks like FAL is a special case of FAN, could you confirm this? I.e., you set the non linearity in FAN to the identity, and this is what causes your improved performance? or are there additional nontrivial changes to the architecture?

Following up, in Table 1/2, is everything between FAN and FAL the same (model architecture, hyperparameters), except the non-linearity?

Will the codebase be fully open sourced upon publication?
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(4) Weak Accept — could be accepted, dependent on rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

While the methodological advancement is mostly incremental (compared to Fourier Analysis Networks (FAN), FAL uses a linear instead of non-linear component), this change seems to be effective as demonstrated on the various benchmark datasets. Also, this seems to be the first demonstration of either FAN or FAL on EEG data, so this part of the evaluation is a contribution nonetheless.

The presentation of the results and discussion is quite weak, and the overall evaluation could be improved in several regards; the paper could be further optimized for clarity.

The authors should confirm that their implementation plus example training scripts will be open sourced so that others can build on the approach.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Review #3

Please describe the contribution of the paper

The paper introduces a novel transformer-based model for EEG emotion recognition that combines frequency-domain analysis and spatial attention. Their key contributions include: (1) the Fourier Analytic Linear (FAL) layer, which decomposes EEG signals into periodic and aperiodic components; (2) the Fourier Adjacency Attention (FAA) scheme, which incorporates a learned adjacency matrix to capture inter-channel relationships and groups attention heads by periodicity; and (3) the resulting Fourier Adjacency Transformer (FAT) architecture that achieves state-of-the-art accuracy on multiple EEG emotion datasets (around 6.5% improvement over previous best).
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1. New architecture: The proposed FAL and FAA modules are fresh contributions to neural network design. They cleverly combined signal processing (Fourier decomposition) and graph-based inductive biases with transformer attention. This innovation directly tackles known problems in EEG modeling (like handling rhythmic vs. non-rhythmic components and incorporating channel connectivity) and is likely transferable to other domains.
2. Good empirical results: The method is empirically very strong, outperforming existing methods by a notable margin on well-known benchmarks (SEED, SEED-IV, DEAP).
3. Clarity and analysis: The authors also analyzed the learned attention structures (via adjacency matrices for periodic/aperiodic components), providing interpretability. The paper is generally well-written and structured
Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1. Notational clarity: There are a lot of notations scattered throughout the paper without proper definition or whose meaning are not entirely clear in the context. For example, in equation 1, 2, and 3, what is W_p, W_n, B_n, what is K, K_p, K_a in equation 4? Some of the notations make sense in the context, but they need to be clarified to make reading the paper smooth for a broader audience. Also, in the results section, it is not stated what the numerical values in the tables represent, I am assuming it is classification accuracy, but it should be clearly stated.
2. Reproducibility concerns: Given the novel layers and mechanisms introduced, reproducing this work could be challenging without access to the authors code or very detailed pseudo-code. Important details such as how the adjacency matrix is initialized, whether it’s fixed or learned per training, and and whether any regularization (e.g., sparsity or symmetry constraints) is applied are not explicitly mentioned in the summary.
3. Minor issues: There are a few minor issues in wording that could be refined. In the abstract, the phrasing “which make accurately emotion classification” is awkward. Also, terms like “Neighbor Attention” and “universal inter-channel correlation patterns” are introduced, which are somewhat intuitive, but ensuring they are precisely defined would help.
Please rate the clarity and organization of this paper

Satisfactory
Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
1. In the results, it would be useful to report not just accuracy but also other metrics or statistical significance. Emotion recognition often also looks at F1-scores or confusion matrices to see if certain emotions are consistently harder. Since the model aims to capture more structure, did it particularly improve classification of certain emotion classes (e.g., maybe it improved distinguishing high-arousal states more)? Any such nuanced insight will be useful.
2. While the authors compare FAT to several graph based models in the results tables, they do not explicitly discuss how their attention based adjacency mechanism differs from or improves upon standard graph neural networks. A brief discussion highlighting the key conceptual or architectural distinctions, such as flexibility, global context modeling, or training stability, would help clarify the advantages of FAT over GNNs.
Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

(5) Accept — should be accepted, independent of rebuttal
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

This paper provides a highly novel and impactful contribution, with clear evidence of superiority over existing methods. The methodology is creative and well-substantiated by results. The potential impact on the field (both technical and for EEG-based applications) is significant. The minor issues noted do not detract from the overall quality; they can be addressed easily.
Reviewer confidence

Confident but not absolutely certain (3)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Author Feedback

We sincerely thank the reviewers and the Area Chair for their thorough review and insightful comments on our manuscript. We are deeply honored by your recognition of our work. We would like to take this opportunity to address the concerns and suggestions raised.

The reviewers’ concerns can be broadly categorized into three main areas, which we will address as follows:

Firstly, regarding further theoretical investigation and clinical analysis of our proposed method: Due to the page limitations of the MICCAI conference, we were unable to include more extensive experiments, explorations, and analyses within this manuscript. The primary focus of this paper was to thoroughly validate the performance of our proposed Fourier Adjacent Attention (FAA) and Fourier Adjacency Transformer (FAT) for EEG-based emotion recognition. We have plans for subsequent in-depth research in more areas, which will be presented in future work.

Secondly, concerning the manuscript’s textual clarity and notation: We will carefully revise and optimize the manuscript based on the specific suggestions provided by the reviewers. Furthermore, regarding the structural relationship between FAL and FAN, as illustrated in Figure 1 and detailed in the experimental section of our paper, the key distinction lies in the activation function applied to the aperiodic component; FAL utilizes an Identity Function to preserve linearity, whereas FAN employs a nonlinear activation. All other hyperparameters were kept identical during model construction and throughout all experiments.

Lastly, and importantly, regarding the reproducibility of this work: We sincerely apologize for our oversight in not explicitly stating the open-source availability of our code in the manuscript. In fact, we have already established an open-source repository for this project and have made a portion of the code publicly available. We are committed to open-sourcing the entirety of the project code in the near future, and we hope this will be a valuable contribution to the EEG decoding community.

We once again express our sincere gratitude to the reviewers for their meticulous work and invaluable advice.

Meta-Review

Meta-review #1

Your recommendation

Provisional Accept
If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

N/A

back to top

A novel Fourier Adjacency Transformer for advanced EEG emotion recognition

Author(s):