Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

High-performance methods for automated detection of epileptic stereo-electroencephalography (SEEG) have important clinical research implications, improving the diagnostic efficiency and reducing physician burden. However, few studies have been able to consider the process of seizure propagation, thus failing to fully capture the deep representations and variations of SEEG in the temporal, spatial, and spectral domains. In this paper, we construct a novel long-term SEEG seizure dataset (LTSZ dataset), and propose channel embedding temporal-spatial-spectral transformer (CE-TSS-Transformer) framework. Firstly, we design channel embedding module to reduce feature dimensions and adaptively construct optimal representation for subsequent analysis. Secondly, we integrate unified multi-scale temporal-spatial-spectral analysis to capture multi-level, multi-domain deep features. Finally, we utilize the transformer encoder to learn the global relevance of features, enhancing the network’s ability to express SEEG features. Experimental results demonstrate state-of-the-art detection performance on the LTSZ dataset, achieving sensitivity, specificity, and accuracy of 99.48%, 99.80%, and 99.48%, respectively. Furthermore, we validate the scalability of the proposed framework on two public datasets of different signal sources, demonstrating the power of the CE-TSS-Transformer framework for capturing diverse temporal-spatial-spectral patterns in seizure detection. The code is available at https://github.com/lizhuoyi-eve/CE-TSS-Transformer.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1378_paper.pdf

SharedIt Link: https://rdcu.be/dV58m

SpringerLink (DOI): https://doi.org/10.1007/978-3-031-72120-5_24

Supplementary Material: N/A

Link to the Code Repository

https://github.com/lizhuoyi-eve/CE-TSS-Transformer

Link to the Dataset(s)

N/A

BibTex

@InProceedings{Li_Epileptic_MICCAI2024,
        author = { Li, Zhuoyi and Li, Wenjun and Zhu, Ning and Han, Junwei and Liu, Tianming and Chen, Beibei and Yan, Zhiqiang and Zhang, Tuo},
        title = { { Epileptic Seizure Detection in SEEG Signals using a Unified Multi-scale Temporal-Spatial-Spectral Transformer Model } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
        year = {2024},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15011},
        month = {October},
        page = {254 -- 264}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose CE-TSS-Transformer, a channel embedding multi-scale temporal-spatial-spectral transformer model, which consists of a channel embedding module, a multi-scale TSS convolution network, a TSS-Transformer network, and a classification network. This model captures the deep representation and variation of SEEG in temporal, spatial, and spectral domains, and has achieved advanced results on three public datasets of various neural activity modalities.

  • Please list the main strengths of the paper; you should write about 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 author used the channel-embedding module to reduce dimensionality and constructed the optimal representation of SEEG. A multi-scale temporal-spatial-spectral analysis was conducted on the optimal representation of SEEG to capture multi-level, multi-domain features. The authors also introduced a Transformer encoder to learn the global dependencies of features. The proposed model has a certain degree of innovation. 2. The authors tested the performance of the proposed method on other neural activity modalities, i.e., deep, strip electrodes and EEG signals.
  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. The model structure is not well presented, especially the wavelet convolutional layer in multi-scale spectral analysis. 2. The model is trained in a patient-dependent way and validated in a leave-one-seizure-out fashion. The generalizability of the proposed model is not clear. I suspect there is over-fitting. 3. Ablation study is insufficient i.e., the necessity of the temporal, spatial, and spectral analysis is not evaluated. 4. Discussions about other methods for extracting multi-domain features or multi-scale features are missing.
  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    No.

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
    1. The generalizability of the model should be further evaluated, for instance, in a cross-patient setup.
    2. Ablation study about the features of each domain should be provided to demonstrate the necessity of multi-scale temporal analysis, spatial analysis and spectral analysis, respectively.
    3. Discussion of results is weak.
  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Lack of evaluation of generalizability, lack of ablation study, lack of deep and insightful discussion.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Reject — should be rejected, independent of rebuttal (2)

  • [Post rebuttal] Please justify your decision
    1. The authors provided accuracy for cross-subject performance. However, due to imbalance in dataset, accuracy cannot objectively indicate the model performance. I still have concerns about the model generalizability. 2. Since new results are not allowed, ablation study is still a weakness. Therefore, I think the manuscript should be rejected.



Review #2

  • Please describe the contribution of the paper

    The author proposed the CE-TSS-Transformer for epileptic seizure detection. And it outperforms the previous methods.

  • Please list the main strengths of the paper; you should write about 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 paper is well-organized and the final performance looks good.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
    1. The ablation study to address the benefits of using multi-scale time/spectral/spatial is not sufficient. It would be great if the author could provide the experiments comparing with the “single-scale” version of the model.
    2. The description of CE module is not clear. In the Fig.1, are those ‘Conv’ layer represent the CE module? And where is the embeding operation in the CE module as shown in the Fig.1? Are you using the ‘Linear’ layer for channel embedding?
    3. For those convolution layer, such as ‘134’ and ‘138’, what are ‘4’ and ‘8’ represents? stride or dilation rate? and what the ‘X’ meaning in the ‘13X’?
    4. The explanation of wavelet convolution is confusing. The X_CE is not a 1-D signal, the equation 2 is confused which dimension to do this operation. And the author mention R is selected as 2 and 8, when R=2, in the equation (2), it will have negative index, how it could work?
    5. It is also valuable to do ablation studies of wavelet types and wavelet levels.
  • Please rate the clarity and organization of this paper

    Very 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 provide sufficient information for reproducibility.

  • Do you have any additional comments regarding the paper’s reproducibility?

    No

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    See in weakness.

  • 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

    Weak Reject — could be rejected, dependent on rebuttal (3)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Though the paper is well organized, the ablation study is not sufficient.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • [Post rebuttal] Please justify your decision

    The author address some of my concerns, however, the weaknesses like imbalance in dataset as well as the ablation study are still there. So I raise my score to weak accept.



Review #3

  • Please describe the contribution of the paper

    The author constructed a long-term SEEG epilepsy dataset and proposed a high-precision monitoring method for epilepsy in this article.

  • Please list the main strengths of the paper; you should write about 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.

    Advantages:

    1.The author constructed a new dataset, which we believe required tremendous effort. Could you please confirm if this dataset will be made publicly available in the future? 2.The author proposed a high-precision multiscale algorithm. 3.This study involved a significant amount of work.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    Disadvantages:

    1.We believe the author should analyze the differences between the constructed dataset and the Bonn Dataset and CHB-MIT Dataset to highlight the contribution of this article. If it is a homogeneous dataset, then it may not significantly advance the field. 2.In the performance comparison experiments of the LTSZ dataset, the contrast models are not strong enough to verify if they provide the best performance. It would be beneficial to include high-precision time series prediction models such as Informer as comparative experiments. 3.Although the proposed model outperforms the Transformer by more than 0.1 in performance, the Transformer’s performance has already reached 0.9802, which does not show a significant difference in practical scenarios. 4.(***** This is the most important issue we are concerned about) After the performance reaches 0.95, we believe that the complexity of the model is a very important reference indicator. The three multiscale approaches proposed in this article will increase the complexity of the model, at least the complexity will be higher than that of Transformer. In this scenario, Transformer might be a better choice. We believe there needs to be a discussion on the complexity of the model. 6.On the sixth page, we did not find any differences between the formulas in TSS-Transformer Net and Transformer.

  • Please rate the clarity and organization of this paper

    Very 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    The model is not complex and is highly reproducible, but will the dataset be made publicly available in the future?

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    1.We believe the author should analyze the differences between the constructed dataset and the Bonn Dataset and CHB-MIT Dataset to highlight the contribution of this article. If it is a homogeneous dataset, then it may not significantly advance the field. 2.In the performance comparison experiments of the LTSZ dataset, the contrast models are not strong enough to verify if they provide the best performance. It would be beneficial to include high-precision time series prediction models such as Informer as comparative experiments. 3.Although the proposed model outperforms the Transformer by more than 0.1 in performance, the Transformer’s performance has already reached 0.9802, which does not show a significant difference in practical scenarios. 4.(***** This is the most important issue we are concerned about) After the performance reaches 0.95, we believe that the complexity of the model is a very important reference indicator. The three multiscale approaches proposed in this article will increase the complexity of the model, at least the complexity will be higher than that of Transformer. In this scenario, Transformer might be a better choice. We believe there needs to be a discussion on the complexity of the model. 6.On the sixth page, we did not find any differences between the formulas in TSS-Transformer Net and Transformer.

  • 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

    Weak Accept — could be accepted, dependent on rebuttal (4)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The rationality of experiments, the innovation of the model, and the clinical application in medicine are the focal points of my consideration.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    Accept — should be accepted, independent of rebuttal (5)

  • [Post rebuttal] Please justify your decision

    Thank you very much to the author for partially addressing my concerns.



Review #4

  • Please describe the contribution of the paper

    The paper explores the multi-scale temporal, spatial, and spectral analyses of long-term epilepsy data for epileptic seizure detection in SEEG signals, which were ignored by existing studies. The authors proposed a new dataset and new approaches for their purpose, and the experimental results have shown their usefulness.

  • Please list the main strengths of the paper; you should write about 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 long-term epilepsy data has seldom been explored by existing studies.

    (2) A novel long-term seizure SEEG dataset was proposed which I suppose would be useful for subsequent researchers working in this area.

    (3) The experimental results have shown the usefulness of their proposed methods.

  • Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.

    The Ablation Study demonstrated the superiority of the proposed technique over some of the commonly used baselines, but did not analyze the effectiveness of each of the important components inside the proposed technique.

  • 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.

  • Do you have any additional comments regarding the paper’s reproducibility?

    N/A

  • Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html

    It would be better to add ablation studies analyzing the effectiveness of each of the important components inside the proposed technique.

  • 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

    Accept — should be accepted, independent of rebuttal (5)

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This manuscript presents an aspect that seems to rarely be addressed in previous work, namely the analyses of long-term epilepsy data for epileptic seizure detection in SEEG signals. Both new datasets and new methods that have been experimentally proven to be effective are proposed for this aspect. In addition, the experiments in this manuscript use publicly available datasets, which enhances the persuasiveness. However, it would be better to add ablation studies analyzing the effectiveness of each of the important components inside the proposed technique.

  • Reviewer confidence

    Not confident (1)

  • [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

    N/A

  • [Post rebuttal] Please justify your decision

    N/A



Author Feedback

We thank reviewers for their comments. Replies are itemized as follows:

Q1(R1&R6):Clarify the wavelet convolution layers in multi-scale spectral analysis. In Eq2, the 1-D signal of length N from each channel in x_CE is taken as input, resulting in x_p after periodic padding. Wavelet decomposition occurs simultaneously across all channels C. We will add subscripts to x_CE for clarity. When R=2, the starting index is N, and the end index is negative. Periodic padding ensures indices exceeding the signal length wrap around to the start.

Q2(R1&R3&R6):Explanation of ablation studies. (1)Multi-scale division is based on the 6 SEEG rhythms. Since single-scale can’t fully capture brain activity, features across all rhythms are necessary. Spectral scales align with temporal and spatial domains. Analyzing scales 1 to 6, we found multi-scale increased temporal, spatial, and spectral accuracies by 15.4%, 13.9%, and 12.9% compared to single-scale, confirming the need for multi-scale feature extraction. (2)Many studies show Db4 excels in processing SEEGs, so we chose it to extract rhythms. We selected wavelet types from db_n, sym_n, and coif_n(n∈{2, 3, 4, 5}), at levels from 1 to 6. Accuracy increases with more levels, with db outperforming sym and coif. At 6 levels, Db4 achieved the best accuracy of 99.89%, confirming our conclusion.

Q3(R1):Discussion of model generalization. We used regularization to avoid overfitting. To evaluate generalization, we conducted cross-patient study with 4 patients for training and 1 patient for testing, performing 5-fold cross-validation. The accuracies for each patient are 1, 0.991, 1, 1, 0.998, averaging of 0.994, confirming the model’s good generalization.

Q4(R1):Discussion of other methods for extracting multi-domain or multi-scale features. Discussion of the results. (1)As stated in the introduction, previous studies overlooked seizure propagation and its multi-domain, multi-scale variations, with few related methods. Existing multi-scale RBF and TripleGAN improve temporal-spectral extraction but ignore spatial propagation and have unstable training. Our method deeply mines seizure features across domains and scales, effectively avoiding these issues. (2)We discussed results immediately after each experiment to clearly demonstrate our model’s performance. We can move them to the discussion section and extend them.

Q5(R6):Clarify the CE module. The part from ‘input SEEGs’ to ‘dynamic subband SEEGs’ is the CE. The embedding of CE consists of a series of ‘Conv’ layers(see in orange in Fig1), which compress and expand channels. No ‘Linear’ layers are used. In Fig1, ‘134’ and ‘138’ denote dilation rates of 4 and 8, while ‘X’ in ‘13X’ denotes the number of SEEG contacts for different patients.

Q6(R1&R3&R5&R6):Publicly available data and code. We will public the code on GitHub and anonymize the LTSZ dataset. If needed, please contact the authors. We will provide necessary data support.

Q7(R5):Analyze the differences between datasets. As stated in the paper, our LTSZ dataset differs from the Bonn and CHB-MIT datasets in neural activity patterns. LTSZ comprises SEEG, CHB-MIT comprises EEG, and Bonn includes subsets C&D with deep electrodes, E with strip electrodes.

Q8(R5):Comparison experiment. When using the time series prediction model Informer for epilepsy detection, the training speed is slow, and the final accuracy is 0.948, lower than our method’s performance.

Q9(R5):Differences with Transformer results. Discussion of model complexity. Considering seizure propagation, the multi-scale, multi-domain captures SEEG patterns better. While TSS-Transformer is slightly more complex than Transformer, it reached 100% detection for all patients except ID2. Our method also extends to epilepsy prediction and localization studies.

Q10(R5):Differences in formulas in TSS-Transformer and Transformer. The TSS-Transformer is formally identical to the Transformer, with the name clarifying module connectivity.



Meta-Review

Meta-review #1

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The reviewers largely appreciated that novelty and importance of the proposed methodology. The rebuttal does a good job of addressing many of the stated concerns.

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    The reviewers largely appreciated that novelty and importance of the proposed methodology. The rebuttal does a good job of addressing many of the stated concerns.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A

  • What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).

    N/A



back to top