List of Papers Browse by Subject Areas Author List
Abstract
Emotion recognition using electroencephalogram (EEG) remains challenging in cross-subject settings because EEG patterns vary across subjects and frequency bands. Most existing methods rely on manually designed backbones, which may not fit the temporal characteristics of all EEG bands. This paper proposes a frequency-aware neural architecture search framework with cross-band fusion for EEG emotion recognition. For each frequency band, we build a searchable directed acyclic graph backbone and optimize its architecture independently. Besides topology search, the temporal configuration inside each node is included in the search space, enabling joint optimization of graph topology and node-level temporal modeling. Each node uses a Bidirectional Mamba module to model past and future context in offline EEG sequences. Cross-subject experiments on DEAP and DREAMER show competitive and stable performance, supporting the use of frequency-aware architecture search for EEG emotion recognition. Code is available at https://github.com/yoghurt125/FANS-BiMamba-Emotion.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0594_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yoghurt125/FANS-BiMamba-Emotion
Link to the Dataset(s)
DEAP dataset: http://www.eecs.qmul.ac.uk/mmv/datasets/deap/
DREAMER dataset: https://zenodo.org/records/546113
BibTex
@InProceedings{XuSha_FrequencyAware_MICCAI2026,
author = { Xu, Shasha AND Zhang, Rongye AND Hao, Yanrong AND Bian, Jing AND Cao, Rui AND Wen, Xin},
title = { { Frequency-Aware Neural Architecture Search with Bidirectional Mamba for EEG Emotion Recognition } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
（1）The paper’s central contribution is a frequency-aware neural architecture search framework for EEG emotion recognition, where the θ, α, β, and γ bands are assigned independently searched architectures instead of sharing one manually designed backbone. This is important because the paper explicitly argues that different EEG bands exhibit heterogeneous temporal characteristics, and a single hand-crafted architecture may not adapt well to all of them. （2）Unlike many DAG-based NAS methods that mainly search edge connections or operation choices, this work explicitly includes the internal temporal modeling configuration of each node in the search space, such as state dimension, convolutional receptive field, expansion ratio, and fusion strategy for Bidirectional Mamba. As a result, the method jointly optimizes global topology and node-level functionality, which substantially increases the expressive power of the searched architecture. （3）The paper also contributes by embedding Bidirectional Mamba into the searchable temporal backbone rather than treating Mamba as a fixed architecture. This allows each node to capture both past dependencies and future context, which is particularly suitable for offline EEG emotion recognition where the full temporal window is available during inference. （4）The work further contributes an empirical validation on the DEAP and DREAMER datasets under leave-one-subject-out cross-validation. It reports 71.41%/72.50% accuracy for valence/arousal on DEAP and 70.76% valence accuracy on DREAMER, and supports the proposed design with ablations comparing random search, fixed architectures, and alternative node operation types.
- 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）A major strength of the paper is that it elevates frequency-band heterogeneity from a preprocessing issue to an architectural design principle. Many EEG methods decompose signals into multiple bands but still process all bands with the same backbone, whereas this work explicitly argues that each band should discover its own optimal architecture through independent NAS. This is an interesting and meaningful design choice, and the distinct DAGs found for different bands provide empirical support for that claim. （2）Another notable strength is the finer-grained NAS formulation: the paper does not restrict search to graph topology or edge operations, but also searches the internal configuration of each node. This effectively extends the search target from “which graph to use” to “which graph and how each node should temporally model the signal,” which is particularly relevant for EEG because temporal dynamics vary substantially across bands and tasks. （3）The experimental setup is reasonably solid and transparent. The paper uses leave-one-subject-out cross-validation for cross-subject evaluation on two public datasets, reports mean performance and some standard deviations, and includes ablations on search strategy and node operation types. This makes the empirical section more convincing than a paper that only reports final benchmark numbers without analyzing where the gains come from.
- 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）The novelty is mainly integrative rather than foundational. Most core ingredients already exist in prior work: reinforcement-learning NAS goes back to Zoph and Le (2017), differentiable/platform-aware NAS is well established, NAS has already been applied to EEG emotion recognition by Li et al. (2022) and Wu et al. (2023), and Mamba/bidirectional EEG Mamba ideas are also present in Gu and Dao (2024) and Gui et al. (2024). Therefore, the paper is best understood as a structured integration of band-specific NAS, DAG search, Bidirectional Mamba, and cross-band fusion, rather than a fundamentally new learning principle. （2）Although the paper strongly motivates band-specific architectures, it does not provide the most decisive ablation: a direct comparison between using one shared searched architecture across all bands and using independently searched architectures for each band. As a result, the paper shows that the design is plausible, but it does not fully isolate whether the gains truly come from frequency-specific specialization rather than from a generally larger or more flexible overall search space. （3）While the baselines include several Transformer- and Mamba-related methods, the comparison is still not fully comprehensive. In particular, the DREAMER table mixes Transformer and miscellaneous methods without presenting a clearly unified set of strong cross-subject baselines under exactly matched preprocessing, segmentation, and evaluation conditions. This means the reported gains are informative, but the fairness and completeness of the benchmarking remain somewhat uncertain. （4）The method does improve performance, but the margin is not always large, and the statistical evidence could be stronger. For instance, in the NAS ablation on DEAP, 71.41% versus 70.54% over random search is a real gain but not a dramatic one; similarly, the advantage over strong single-operation nodes such as TCN (71.33%) and LSTM (71.09%) is relatively modest. The results therefore suggest stable incremental improvement rather than a decisive leap, and stronger statistical testing would make the claim more convincing.
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
N/A
- 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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper proposes a frequency-aware neural architecture search framework for cross-subject EEG emotion recognition. The central idea is that different EEG frequency bands exhibit distinct temporal characteristics, so they should not necessarily share the same fixed network architecture. Instead, the method performs independent architecture search for the θ, α, β, and γ bands, allowing each band to discover its own optimal backbone. To achieve this, the temporal backbone of each band is represented as a directed acyclic graph, and the search process jointly optimizes not only the overall topology but also the internal temporal modeling configuration of each node. Each node employs a Bidirectional Mamba state space model to capture both historical dependencies and future context, and the resulting band-specific representations are aggregated through a cross-band attention fusion module for final emotion classification.
The method is evaluated on the public DEAP and DREAMER datasets under a leave-one-subject-out cross-validation protocol. The reported results show competitive performance, including 71.41% valence accuracy and 72.50% arousal accuracy on DEAP, as well as 70.76% valence accuracy on DREAMER. The paper also includes ablation studies on the effectiveness of NAS and on different node operation types, suggesting that jointly searching network topology and node-level temporal configurations is beneficial. Overall, the main contribution of the work lies in integrating frequency-specific modeling, neural architecture search, Bidirectional Mamba, and cross-band fusion into a unified framework for cross-subject EEG emotion recognition. At the same time, its novelty is better characterized as a systematic integration and extension of existing ideas rather than a fundamentally new paradigm, and the paper leaves room for stronger efficiency analysis and broader baseline comparisons.
- Reviewer confidence
Somewhat confident (2)
- [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
This paper proposes a frequency-aware NAS framework for cross-subject EEG emotion recognition. It performs separate architecture search for different frequency bands, jointly optimizing edge connections and bidirectional Mamba configurations within nodes in a DAG search space, and then combines the resulting representations through cross-band attention fusion for classification.
- 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.The paper wants to addresses two central challenges in cross-subject EEG emotion recognition: substantial inter-subject variability and temporal heterogeneity across frequency bands. Instead of using a single manually designed backbone for all bands, it searches architectures separately for $\theta/\alpha/\beta/\gamma$, which is more consistent with the intuition behind multi-band EEG modeling and has potential value for the community. 2.The method also shows some originality. In addition to searching edge connections and operation types in the DAG, it includes the temporal module configuration inside each node in the search space, making it possible to optimize network topology and node functionality jointly. 3.The experimental results are competitive. The paper reports valence/arousal accuracies of 71.41%/72.50% on DEAP and a valence accuracy of 70.76% on DREAMER. In addition, the searched structures for different frequency bands provide preliminary evidence that different bands may favor different temporal architectures.
- 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.That said, the NAS experiments lack important information needed to assess fairness and reproducibility. The paper does not clearly report the search budget, whether each LOSO fold is searched independently, or the total search cost and final model complexity. As a result, it is difficult to determine whether the reported gains come from a better search strategy or simply from greater search cost. This is a central issue for NAS papers. 2.The comparative experiments are also not fully convincing. Many of the baselines in Tables 1 and 2 appear to be taken directly from prior work rather than reimplemented under a unified experimental protocol. Since EEG emotion recognition is highly sensitive to preprocessing and evaluation details, the current results are more sufficient to show competitiveness than to support a strong claim of superiority over existing methods. This is particularly relevant on DEAP, where the gain over the strongest Mamba baseline is only about 1–2 percentage points, and on DREAMER, where the method does not achieve the best arousal performance. 3.The necessity of Bidirectional Mamba has not been fully proven. Table 4 shows that when a fixed single node type is used, the performance of TCN and LSTM is already very close to the final method. In addition, the claimed advantage of band-independent search still needs direct validation. Currently, there is a lack of the most direct ablation experiment of performance: sharing the architecture across bands vs. independent search for each frequency band. 4.Figure 2 shows the different frequency band DAG structures obtained from a certain search. However, under the LOSO-CV setting, the training/validation split of different folds will change. If the optimal frequency band structures obtained from each fold of search vary greatly, then “band-specific architecture preference” may only be adapted to a specific data partition and may not be a stable cross-subject rule. So, does the DAG architecture obtained from different frequency bands have cross-subject generalization significance?
- 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.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
N/A
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I assigned this paper a Weak Reject because it shows some methodological originality and competitive empirical performance, but the current evidence is still not strong enough to place it clearly above the acceptance threshold.
On the positive side, the paper addresses two genuine challenges in cross-subject EEG emotion recognition: inter-subject variability and temporal differences across frequency bands. The choice to search separate architectures for different bands is well motivated, and the search space goes beyond standard DAG-based NAS by including not only edge operations but also the temporal configuration within each node. The results on DEAP and DREAMER are also competitive, suggesting that the overall framework is promising.
My main concerns lie in the strength of the experimental evidence. For a NAS paper, the lack of detail on the search budget, the per-fold search protocol, the total computational cost, and the final model complexity makes it difficult to judge fairness and reproducibility. The comparison with prior work is also not entirely convincing, since many baselines appear to be taken from the literature rather than reimplemented under a unified setting. In addition, two central claims still need stronger support: the necessity of Bidirectional Mamba and the advantage of band-specific search over a shared architecture. These issues seem potentially addressable in rebuttal, which is why I see the paper as borderline rather than a clear reject.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I still view the paper as borderline. The rebuttal addresses several major concerns and adds useful details on the search protocol, ablation studies, and computational cost, which makes the work more credible. However, much of the key supporting evidence appears only in the rebuttal and has not yet been integrated into the manuscript. The required revisions also go beyond minor clarifications. They would require substantial changes to the description of the NAS protocol, the baseline comparison, the complexity analysis, the core ablations, and the contribution statement. I therefore appreciate the authors’ response and the added evidence, but I still regard the paper as close to the acceptance boundary rather than clearly above it.
Review #3
- Please describe the contribution of the paper
he paper proposes a neural architecture search approach for models analyzing EEG data. Specifically the NAS addresses three limitations of the state of the art relevant in this context. Prior approaches limit search to a relatively small model familiy such as convolutional models, they do not exploit different roles of frequency bands in this type of data, and they change the node connections instead of the role of nodes in directed acyclical graphs used to represent the model during the search. The results suggest that this strategy leads to superior results compared to a range of state of the art methods. While the paper proposes a sophisticated approach and outperforms established methods, the paper could use work in being more easily accessible to non-experts. Currently, it is hard to gain intuition what the search actually finds, and how the resulting models differ from others - e.g., fig.2 shows the resulting models from the proposed method - would it make sense to also show those of others, to understand if the method performs relatively modest fine tuning, or if it find radically different model architectures.
In summary this is a paper well suited for MICCAI, but the authors should improve on understandability and accessibility for non experts.
- 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.
- Novel approach addressing a relevant problem
- The NAS strategy might be applicable to other problems, and I would encourage the authors to reflect on that in the discussion, to make the paper also for non EEG researchers useful
- The results suggest that the method outperforms the state of the art.
- 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.
- Description of method and discussion of results could be improved. For instance the discussion of the intuition behind the improvements would help appreciate the result relevance
- 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 provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- 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/2026/en/REVIEWER-GUIDELINES.html
- Please add plots also of SoA alternatives in Fig2
- In Table 1 and 2 please make clear which models are handcrafted models, vs. which are also NAS approaches.
- 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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting approach, and shows advantages compared with the state of the art. At the same time, clarity could be improved.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I already recommended weak accept before the rebuttal and stay with this assessment
Author Feedback
We thank the reviewers for their comments. We agree that the manuscript should be strengthened in reproducibility, search protocol, baseline transparency, and ablations. Added analyses will be used as supplementary evidence based on existing results, not as new main claims.
Regarding novelty, we acknowledge that the contribution was not sufficiently clear. RL based NAS, DAG search, Mamba, and bidirectional modeling are established techniques. Our contribution is to incorporate frequency band heterogeneity into architecture search for cross subject EEG emotion recognition. Rather than using one shared manual backbone, the framework searches a band specific DAG topology and node level Bidirectional Mamba configuration for each band. We will refine the novelty statement and emphasize this band aware NAS framework.
Regarding NAS fairness and reproducibility, architecture search is conducted once on a fixed subject level training and validation split. The selected architecture is then fixed and retrained from scratch under LOSO CV, so all folds use the same searched architecture rather than fold specific ones. This avoids fold specific tuning while keeping the search tractable and reproducible. Across multiple random seeds, each search run evaluated 200 candidates, each trained for at most 80 epochs with early stopping. The search runtime was 2.79 ± 0.36 GPU hours per run. The final models contained 1.80 ± 0.28M parameters and required 0.94 ± 0.18 ms per sample. We will report FLOPs and memory usage.
Regarding baselines, EEG emotion recognition is sensitive to preprocessing, segmentation, and training details, so cross paper comparisons should be cautious. All compared methods follow LOSO cross subject evaluation. For methods with public and stable code, reimplementation was attempted. For methods without code or with unreliable reproducibility, reported results were used. If reproduced results were lower, reported values were kept to avoid underestimating competing methods. In Tables 1 and 2, we will mark each result as reported or reimplemented and clarify the method category. The conclusion will be revised to competitive and stable performance, rather than strict superiority under a unified benchmark.
Regarding the shared architecture ablation, we added a direct comparison under the same data version, subject level search split, search space, and LOSO protocol. Independent band specific search improved accuracy by 0.93 percentage points over the shared searched architecture. It also showed lower final complexity, with fewer parameters, lower MACs, and lower inference latency. For search runtime, the shared searched architecture required about 3.3 times the search time of independent band specific search. These results suggest benefits in both performance and efficiency, rather than only greater architectural flexibility. We will report GPU hours, complexity metrics, and paired statistical results.
Regarding Bidirectional Mamba, we agree that TCN and LSTM perform close to the final method. Thus, we will not claim that BiMamba has a decisive advantage based only on mean accuracy. Our focus is to clarify that, when combined with searchable DAG topology and band specific modeling, BiMamba provides stable bidirectional temporal modeling under acceptable complexity. We will add parameters, FLOPs, inference time, and memory usage for different node types to examine whether the differences come from structural design rather than model capacity or computational cost.
Finally, regarding Fig.2 and readability, we will clarify that Fig. 2 shows representative structures from repeated searches, not architectures customized for individual test subjects. The exact connections may vary, while the overall band specific trends are similar. We will interpret them as empirical evidence of band specific architectural preferences, not universal neurophysiological rules. We will also explain DAG nodes, edges, and search results more intuitively.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper received 3 borderline with two reviewers suggesting weak accept and 1 weak reject. All reviewers identify originality in the methodology and concrete validation as major strengths of the work. Crucially though key weaknesses are identified that need clarification. In summary, authors are instructed to address:
1) There is agreement that reproducibility is not straightforward. This is a major concern.
2) Major methodological questinos raised by Rev #1 and #2
3) Clarifications on experimental setup, discussion and analysis of results as raised by all reviewers
4) Suggestions for ablation studies and statistical testing from Rev #1
The suggested additional experimentation should be attempted on the already available set of results. No new main results should be produced.
Me recommendation is for the paper to be invited for rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal has reviewed by 2/3 reviewers who elevated their scores to accept. Depsite this I am very sceptical for this paper. The rebuttal does clarify many important elements however these must be integrated and the paper considerably refocused (aka rewritten). There is also no comittment of public release of code to enhance reproducability, which was a major concern but not addressed in the rebuttal.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers agree that the paper presents a relevant and technically sound frequency-aware NAS framework for cross-subject EEG emotion recognition. Its main strengths are the band-specific architecture search, joint optimization of DAG topology and node-level temporal modeling, and competitive results on DEAP and DREAMER. Although concerns remain about search cost, baseline fairness, key ablations, and clarity, the rebuttal addressed major issues and changed the borderline negative opinion to Accept. Overall, the paper is solid and relevant to MICCAI, so I recommend Accept, with final revisions incorporating the rebuttal clarifications.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal addresses the reviewers’ main concerns. The authors provide a clearer novelty statement, additional details on the NAS protocol and computational cost, and a better interpretation of the BiMamba component and baseline comparisons. While reproducibility remains somewhat complex and the final manuscript should explicitly clarify whether the architecture-search split is independent of LOSO test subjects. I would recommend acceptance.