Deep learning based neural decoding from stereotactic electroencephalography (sEEG) would likely benefit from scaling up both dataset and model size. To achieve this, combining data across multiple subjects is crucial. However, in sEEG cohorts, each subject has a variable number of electrodes placed at distinct locations in their brain, solely based on clinical needs. Such heterogeneity in electrode number/placement poses a significant challenge for data integration, since there is no clear correspondence of the neural activity recorded at distinct sites between individuals. Here we introduce seegnificant: a training framework and architecture that can be used to decode behavior across subjects using sEEG data. We tokenize the neural activity within electrodes using convolutions and extract long-term temporal dependencies between tokens using self-attention in the time dimension. The 3D location of each electrode is then mixed with the tokens, followed by another self-attention in the electrode dimension to extract effective spatiotemporal neural representations. Subject-specific heads are then used for downstream decoding tasks. Using this approach, we construct a multi-subject model trained on the combined data from 21 subjects performing a behavioral task. We demonstrate that our model is able to decode the trial-wise response time of the subjects during the behavioral task solely from neural data. We also show that the neural representations learned by pretraining our model across individuals can be transferred in a few-shot manner to new subjects. This work introduces a scalable approach towards sEEG data integration for multi-subject model training, paving the way for cross-subject generalization for sEEG decoding.
Paper
Similar papers
Peer review
Summary
The work proposes a transformer architecture for decoding behavior using sEEG neural recording which can account for inter-subject variability that exist due to different placements of sEEG electrodes, different SNRs, inherent biological differences etc. The ability of the proposed transformer architecture to deal with inter-subject variability allows the transformer architecture to leverage a much larger multi-subject dataset as compare to existing works which rely on subject-specific models. The proposed architecture consists of six main components: (i) a CNN temporal tokenizer; (ii) A temporal attention mechanism; (iii) A hand-designed RBF-based spatial encoding of different electrode locations; (iv) A spatial attention mechanism; (iv) a feed-forward MLP to extract small dimension neural representations; and (iv) a multi-head regression layer to deal with inherent patient-to-patient variability. The authors test their proposed architecture on a dataset collected from 21 subjects performing a visual color-change task, and show that their proposed architecture significantly boosts its decoding performance when using multi-subject datasets as compared to single-subject dataset.
Strengths
- The paper is well-written and easy to follow. - The paper addresses an important issue of data heterogeneity in sEEG recordings among participants which makes constructing large dataset challenging as data from multiple participants cannot be easily pooled together. This particularly affects the applicability of deep learning models as they typically rely on large datasets to get low generalization error. - The proposed architecture is clearly defined and uses a multi-head regression layer with a shared trunk (similar to some multi-task architecture) to model inter-patient variability. The authors also employ spatial and temporal attention layers to model temporal dependence in the neural activity and the spatial dependence between different electrodes, respectively. and - The experimental details are clearly provided and the results shown in the work clearly show that training transformers on multi-subject dataset significantly boosts their decoding performance. - The ablation study is insightful and provides important insights into the functioning of different components of the proposed architecture.
Weaknesses
1. A main weakness of this paper is the spatial encoding (which is framed as a main contribution of the work) only seems to marginally improve the performance of the transformer architecture. The results shown in Section 4.5 only show that average decrease of 0.02 in $R^2$ coefficient across the subjects which is well within the standard error of 0.05 (line 279) reported by authors for the average $R^2$ coefficient with the transformer model incorporating spatial encoding. I am surprised why the authors do not use this result to conclude that the proposed spatial encoding does not have a significant effect on model's accuracy and, consequently, the heterogeneity in the *actual* positions of sEEG electrode does not seem to impede the training of multi-subject models. Looking at authors' result of ablation study in A.3.1, it seems more plausible that attention mechanism for space (without which the $R^2$ coefficient decreases by a more significant amount of ~0.1, figure 6) is the main mechanism by which the proposed architecture deals with electrode heterogeneity. 2. Furthermore, the ablation study detailed in A.3.1 seems to suggest that it is the multi-head regression component of the proposed architecture is the main component enabling the better performance of multi-subject model. I am surprised as to why this observation is not mentioned in the main manuscript. 3. Due to the marginal performance boost offered by positional encoding, the argument that heterogeneity of electrode placement needs to be explicitly accounted for by tokenizing space and time separately seems weak (lines 161-164). Consequently, the separation of time and space attention mechanism does not seem to be adequately justified. I think comparison against 2-D attention mechanism will benefit the study. 4. Comparison against other methodology is limited. Authors only compare their architecture on single-subject dataset with multi-subject dataset and show a significant increase in the performance of their transformer architecture when training with the larger multi-subject data. Naively training transformer architecture without heavy regularization on small single subject dataset seems to provide poor performance as already discussed in Ye & Pandarinath'21. Hence, the gains of using multi-subject dataset reported in this study might be artificially larger than the ones obtained with proper regularization (such as dropout, $L_2,$ and $L_1$ regularization) of the transformer architecture. Furthermore, authors should also use some classical decoding models such as (PCA with linear regression, Unscented Kalman Filters, see Xu et al. JNE'22) to give a sense of how much gain is obtained in the first place by using the more complicated transformers architecture and provide a baseline accuracy for comparison. Authors should also compare against existing techniques such as SEEG-Net (Wang et al. Computers in Biology and Medicine'22) which used domain generalization techniques for dealing with heterogeneous sEEG data distribution among different subjects. As a side note, I think the work would benefit from a discussion on multi-task literature, as the multi-head regression structure is quite similar to many multi-task architecture proposed earlier (shared trunk architecture, see the review Crawshaw'20 on arvix). Minor: - Line 210: "long-*range* dependency" might be more appropriate than "long-*term* dependency" as the authors are talking about variables representing space. - Please provide the test/train/validation splits used in this work. It would facilitate in understanding the reported test $R^2$ coefficient.
Questions
1. Line 230: What is C? Should it be E? 2. Section 3.2.2: Spatial Positional Encoding: A couple of questions: - I am assuming that an MRI for each subject in the study was available, from which the MNI coordinates of the sEEG electrodes were calculated. In the spirit of being fully data-driven, why not use the whole MRI (with the sEEG shanks) for encoding the spatial latents? A simple CNN or Graph Convolutional Net could be used to directly encode the spatial information into a latent which can be used instead of the hand-designed coordinate-based spatial feature used in this work. A minor issue might be that the spatial latents might also contain information about the electrodes discarded in the preprocessing but that can possibly be remedied by appropriate masking or removing the sEEG coordinated from the MRI while pre-processing (whichever is easier). A possible advantage of directly using the MRI data would be that it would also account for differences in the brain structure along with the difference in sEEG probe placement. - I do not understand the point of using $m$-different 1-D RBFs for modeling the positional encoding. Is the goal to model positional encodings at different spatial scales. Why is information at different spatial scales relevant? How are the different $\sigma_j$ chosen (are they learnt from the data or chosen as hyperparameters)? The choice seems somewhat arbitrary, especially since the gain in the model's performance by using the spatial encoding is marginal (a 0.02 increase in R$^2$) as reported by the authors in Section 4.3. - Is the rationale for using 1-D RBFs, i.e., the RBF for each coordinate (eq (4)) instead of using 3-D RBFs (e.g., something like $C\exp(-D((x-\mu_x)^2+(y-\mu_y)^2+(z-\mu_z)^2$)) is to also incorporate the information of the direction of the coordinate. - Lines 205-207: The authors state that the projected positional encodings are added to the latents. I am having trouble fully understanding what does this operation entail? Does this mean that the projected positional encodings, denoted by $p\in\mathbb{R}^K$, and the latents, denoted as $z_{int}^{3} = z_{int}[i,j,:]$ (using the pythonic notation) for some $0 \leq i \leq E-1$ and $0 \leq j \leq T-1$, are added to produce the output $o\in\mathbb{R}^{E\times T\times K}$ of the operation in the following manner: $o^{3}=z_{int}^3+p$ followed by appropriate stacking. If that is the case, then what are the relative magnitudes of $z_{int}^{3}$ and $p$? Since $p$ is calculated using the 1-D RBFs, depending upon the magnitude of $\sigma_j$'s, the corresponding projected $p$ could be really numerically small, and the corresponding output $o^3=z_{int}^3+p$ might be dominated by the term $z_{int}^3$. I think checking the relative magnitude of $z_{int}^3$ and $p$, and ensuring that they roughly have comparable magnitudes is important to ensure that the positional encoding impact the final outcome. 3. Why is Huber loss used for training multi-subject model whereas MSE loss used for training single-subject loss? Ideally both multi-subject and single-subject models should have been trained using both losses and for each model, the Huber or MSE loss should have been chosen based on validation error. The choice is surprising as I would have expected to use the more robust Huber loss on the single-subject model (which would be more noisy and prone to effects of outliers) and MSE on the multi-subject data where the much larger size of the dataset can regularize the training process. 4. Another point of potential unfair comparison is that single-subject models are trained for much smaller number of total (gradient) descent steps compared to multi-subject models. On average, the multi-subject models have 21x larger dataset compared to a single-subject model. Hence, a single epoch while training multi-subject models will go through 21 x more gradient descent steps compared to an epoch of a single-subject model. So, multi-subject model trains for 21x longer than the single-subject model. I think authors should provide a justification for this large discrepancy in the training periods of multi-subject and single-subject models. For example, try training the multi-subject model for 1000/20 = 50 epochs and see its $R^2$ coefficient or try training a couple of best-performing single-subject model and train them for 21000 epochs and see their performance. **Edit**: This comment is no longer applicable. I missed the part about different batch-sizes.
Rating
5
Confidence
3
Soundness
2
Presentation
4
Contribution
3
Limitations
Yes, the authors provide a discussion on some limitations in the discussion section.
Summary
The authors propose a novel training framework and architecture to predict response time for a color change behavior task from stereotactic electroencephalography (sEEG) data, focusing on integrating data across multiple subjects despite the variability in electrode placement and count. The model tokenizes neural activity within electrodes using convolutions, extracts temporal dependencies with self-attention, incorporates electrode location with a positional encoding scheme followed by another spatial self-attention layer, and a subject-specific prediction layer. The model is trained on data from 21 subjects, using different procedures: single-subject training (baseline), multi-subject training, and multi-subject training plus single-subject finetuning. The proposed method demonstrates an improved ability to decode trial-wise response times compared to the baseline, and transferability of learned representations to new subjects.
Strengths
* **Innovative Framework**: The proposed framework effectively addresses the heterogeneity across subjects, a significant challenge in sEEG data processing, by using electrode placement and subject-specific prediction layers. The ability to pretrain the model on a larger multi-subject dataset and fine-tune it for new subjects with minimal training is a valuable feature, enhancing the model's practicality and applicability. * **Comprehensive Evaluation**: The model's performance is validated using a substantial dataset (21 subjects), showing consistent improvement in performance for most subjects. An ablation study is also performed.
Weaknesses
The paper discusses the impact of the positional decoding in the ablation study, but this is the least impactful structure compared to other layers (Fig 6). It would be beneficial to explore other position decoding mechanisms besides using MNI locations for future work, perhaps based on different brain atlases.
Questions
1. With the number of electrodes differing across subjects, how did you handle the different input sizes? 2. Instead of plotting all the sEEG channels, it would be helpful to show the selected channels actually used in training. 3. What is the (relative) root mean square error for the prediction?
Rating
7
Confidence
4
Soundness
4
Presentation
3
Contribution
3
Limitations
The method could provide valuable insights for various BCI studies if the authors can demonstrate its performance on more complex tasks. In the Discussion, the authors mention that developing a multi-task model is planned for future work.
Summary
This paper presents a framework and architecture for decoding behavior across subjects using stereotactic electroencephalography (sEEG) data, addressing the challenge of electrode variability. By tokenizing neural activity with convolutions and employing self-attention mechanisms along with a positional encoding scheme based on MNI coordinates, the model extracts effective spatiotemporal neural representations. The study demonstrates successful decoding of behavioral response times from 21 subjects' data and shows that the pretrained neural representations can be transferred to new subjects with minimal training data. This work offers a scalable approach for integrating and decoding sEEG data across multiple subjects.
Strengths
1. Multi-Subject Generalization: The framework's ability to generalize across subjects by combining data from multiple individuals and training a unified model is a substantial step forward compared to traditional single-subject approaches. 2. Methodology: The detailed methodology, including signal preprocessing, and bootstrap randomization test for identifying significant electrodes, ensures the robustness of the proposed approach. 3. Clarity: The paper is well-organized and clearly written, making it accessible to readers from both neuroscience and machine learning backgrounds
Weaknesses
Plz go and check questions.
Questions
1. Scale and diversity of the dataset: The paper mentions using data from 21 subjects, totaling 3600 behavioral trials and 100 hours of sEEG recordings. Is this amount of data sufficient to represent the neural activity patterns of all subjects? Could the diversity of the dataset be insufficient to support the model's generalization capabilities? 2. Effectiveness of spatial positional encoding: The paper proposes a spatial positional encoding method based on MNI coordinates to handle electrode placement variability across subjects. Has the effectiveness of this method been thoroughly validated? Are there comparative experiments showing that this encoding method is superior to other possible encoding methods? 3. Individualized decoder heads: The paper mentions that each subject has a personalized task head for downstream decoding tasks. How effective is this approach when dealing with new subjects? Are there detailed experimental results demonstrating the performance of this method on new subjects? 4. Computational complexity and scalability: The paper outlines a complex model architecture involving convolutional tokenization, self-attention in both time and electrode dimensions, and individualized regression heads. What are the computational requirements for training and running this model? Is the approach scalable to larger datasets or real-time applications?
Rating
4
Confidence
4
Soundness
2
Presentation
3
Contribution
2
Limitations
Plz go and check questions.
Thank you for your response and for updating your score! We are very happy that our responses and new experiments addressed at least some of your concerns. Please find our answers to your new comments below: > Thank you for the authors' efforts and detailed responses. However, I still have concerns regarding the scale of the dataset. While the authors mention that recent sEEG datasets contain fewer than 12 subjects, this does not sufficiently justify that 23 subjects are adequate. In related fields, datasets often include 50, 100, or even more subjects to ensure robustness and generalizability. We echo that a larger dataset (50-100 subjects) would provide stronger evidence for the model's generalization capabilities. However, as reviewer o6UF mentioned "gathering this 50-100 sEEG subjects could take years, if not decades, not to mention the additional difficulty of recruiting them for behavioral studies". Importantly, if the scientific community found value in studies containing 12 subjects or less, it is likely that it will find value in our 21-subject study as well. > I acknowledge the challenges of data collection from a medical perspective. However, there are still larger, publicly available sEEG datasets. This is very insightful. _If you could point us towards those datasets, by specifying either the publication with which they are associated or the link to the dataset repository, we would be very thankful_ and we look forward to working with those datasets in the future. > If collecting sEEG data in epilepsy is too difficult, it might be worth considering using other types of brain signals, which could be more suitable for applying machine learning techniques. Thank you for your suggestion. While there are certainly many different types of brain signals that can be analyzed using machine learning techniques, possibly with larger datasets, we believe that there is a lot of value in applying machine learning techniques to sEEG datasets because: 1. sEEG is currently the gold standard invasive neural recording modality used in humans. Therefore, improving neural decoding based on sEEG bring it closer to clinical translation compared to other neural recording modalities which might provide a lot of data from animal models but have very rarely been used in humans due to safety or other concerns (such as microelectrode array recordings for example). **Edit**: We would also like to emphasize that _even if another neural recording modality was used, the only clinical population approved for collecting intracranial neural recordings would be epilepsy patients_. Therefore, the challenges associated with collecting a large dataset would carry over to any other modality as well. 2. there is a lot of scientific value in showing that machine learning tools can be used on smaller datasets, as in numerous fields (healthcare & medicine, astronomy, environmental science, to name a few) collecting large datasets can be extremely costly and time-consuming. Our work is especially valuable in that sense, since we show that combining many small datasets, despite data heterogeneity, can lead to better machine learning models, compared to training on the smaller datasets individually. Thanks for engaging in the discussion! We very much appreaciate your insights and truly believe that our work has become stonger based on your feedback and suggestions.
Summary
The paper presents a novel approach to sEEG decoding. Authors highlight the benefit of using data from various subjects for training. However, due to the nature of the sEEG technique, collection of such data is difficult. Authors provide a new deep learning based decoding approach which utilizes spatial positional encoding (to provide a model with information on electrodes’ locations), temporal and spatial attention, MLP and subject-specific regression heads. Then, authors train their model in multiple frameworks and highlight the efficacy of multi-subject training.
Strengths
The paper highlights a novel approach to sEEG decoding which enables training of their model on multiple subjects. Generalization across subjects is overall a very important and difficult task to achieve in many modalities and applications of neuroimaging. Authors present a method which has a value in itself and provides ideas for future research in this direction. A paper clearly presents the ideas and performed experiments. Visualizations help to understand both the data collection process and the deep learning architecture.
Weaknesses
1. No comparison with other methods during within-subject training was done. Authors only provide the results for their own architecture for within-subject experiments. Therefore, they only show that their multi-subject trained model is stronger compared to their single-subject model. However, it might be possible that some of the other existing single-subject State-of-the-Art models will be more effective than authors’ multi-subject model on authors’ data. The importance of these comparisons are magnified by the fact that authors’ dataset will stay private (therefore only the authors of this paper can provide metrics for other models applied to this dataset); and by the reasonably large size of authors’ model (it might be too big to be efficiently trained on a single person data, while other models, which were developed for single-subject tasks, might be lighter and train better on small datasets). 2. Decoding performance with and without spatial positional encoding seems very similar. For such a difference it is interesting to see if it has a statistical significance.
Questions
1. How many trainable and frozen parameters there were during training and finetuning of the model? 2. How large is the variance in the response time within each subject? 3. Are all K convolution kernels in Convolution tokenizer the same for all electrodes? 4. How much recording hours of data was used overall and on average per subject (only the electrode-hours are provided)? 5. What is the statistical significance of the metrics increase while using positional encoding compared to not using it?
Rating
6
Confidence
4
Soundness
3
Presentation
3
Contribution
1
Limitations
Yes, authors provide a good description of the limitations and address potential topics for future research (larger datasets, self-supervised pre-training).
Thanks for the detailed response and additional analysis. The new experiments substantially enhance the quality of the work.
Thank you for your response. We are happy to read that our new experiments substantially enhanced the quality of the work.
I have read the author rebuttal and increased my overall score to 5. - I still think the biggest weakness of the work is comparison with existing baselines. Authors do not compare against any existing state of the art method, e.g., Ye & Pandarinath'21 or SEEG-NET. - Authors should also not claim that "This work is the first to train a unified, multi-session, multi-subject models for neural decoding based on sEEG", when SEEG-Net (Wang et al. Computers in Biology and Medicine'22) has done it two years earlier. - I think it is a bit of a stretch to claim that $\\|p\\|$ (which is $\sim$**50** times smaller than the standard deviation in $\\|z^3\\|$) is affecting the output $o$. - I do not agree with the authors' assertion that 3-D RBFs require much more parameters. The common RBF kernel is of the form $\exp(-\\|x-x_0\\|^2/\sigma^2)$ which only requires a single parameter $\sigma$, which would result in exactly the same number of parameters. You could potentially increase the numbers of parameters by using the Mahalanobis distance instead of the standard distance but even in that case the number of parameters for each RBF kernel is $6$ instead of $1$ (overall parameters 6x9 = 54), which compared to the number of parameters in the transformer architecture are negligible. Also I do not understand how 1-D RBFs are able to sample a 3-D space using $3m$ points. Note that if I am discretizing each dimensions using $m$ bins, then I have discretized the 3-D space into $m^3$. Consider the example where I discretize each dimension into two bins {0,1}, then I access 8 different bins in 3-D space as {0,0,0}, {0,0,1}, {0,1,0}, {0,1,1}, {1,0,0}, {1,0,1}, {1,1,0}, and {1,1,1}.
Thank you for your response and for updating your score! We are very happy that our responses and new experiments addressed at least some of your concerns. Please find our answers to your new comments below: > * I still think the biggest weakness of the work is comparison with existing baselines. Authors do not compare against any existing state of the art method, e.g., Ye & Pandarinath'21 or SEEG-NET. Based on your initial feedback, we put a lot of effort into comparing our models with baselines. However, you would like us to compare against: 1. Ye and Pandarinath (2021): The NDT model is designed to "transform sequences of binned spiking activity into inferred firing rates" (this is directly quoted from the paper: line 1 of section 2). NDT operates on _single-unit electrophysiology datasets, which can only be recorded with microelectrode arrays_. In this work, we use an sEEG dataset and it is impossible to resolve single units from sEEG recordings. Given the nature of our dataset, it is unclear how we could use NDT on our data. 2. Wang et al. (2022): sEEG-Net is a model designed for pathological activity detection for drug-resistant epilepsy. It accepts univariate timeseries of neural activity from _single_ sEEG electrodes and classifies them as: physiologic, pathological, or artifact (it maps $R^{T} \rightarrow R^{3}$). In contrast, in this work we are dealing with a problem where multi-variate timeseries of neural activity from _multiple_ sEEG electrodes need to be mapped to a response time (map $R^{E \times T} \rightarrow R^{1}$). Therefore, without extensive modificaton, sEEG-Net cannot work on our data. Adding to this, _the code for sEEG-Net is not public, hampering our efforts to reproduce the model._ We hope you understand that we have been unable to provide those comparisons for solid reasons. > Authors should also not claim that "This work is the first to train a unified, multi-session, multi-subject models for neural decoding based on sEEG", when SEEG-Net (Wang et al. Computers in Biology and Medicine'22) has done it two years earlier. We respectfully disagree. We believe our claim is justified because: 1. sEEG-Net is not a neural decoding model, rather a pathological activity detection model. Neural decoding and pathological activity detection are two very different tasks. 2. sEEG-Net operates on single-electrodes (univariate timeseries of single sEEG electrodes) while our models operate on single-subjects (multivariate timeseries of many sEEG electrodes). > * I think it is a bit of a stretch to claim that $||p||$ (which is 50 times smaller than the standard deviation in $||z||^{3}$) is affecting the output $o$. We echo that the contribution of $||p||$ is very small. We will make sure to discuss this limitation in our revised manuscript and emphasize the need to identify better spatial positional encoding schemes, such as encoding whole brain MRI scans or using atlases other than MNI, as you and reviewer o6UF suggested. > * I do not agree with the authors' assertion that 3-D RBFs require much more parameters. The common RBF kernel is of the form $exp(-||x-x_{0}^{2} / \sigma^{2})$ which only requires a single parameter $\sigma$, which would result in exactly the same number of parameters. You could potentially increase the numbers of parameters by using the Mahalanobis distance instead of the standard distance but even in that case the number of parameters for each RBF kernel is 6 instead of 1 (overall parameters 6x9 = 54), which compared to the number of parameters in the transformer architecture are negligible. Also I do not understand how 1-D RBFs are able to sample a 3-D space using $3m$ points. Note that if I am discretizing each dimensions using $m$ bins, then I have discretized the 3-D space into $m^{3}$. Consider the example where I discretize each dimension into two bins {0,1}, then I access 8 different bins in 3-D space as {0,0,0}, {0,0,1}, {0,1,0}, {0,1,1}, {1,0,0}, {1,0,1}, {1,1,0}, and {1,1,1}. We agree that the number of parameters saved by using a 1D RBFs instead of a 3D RBFs is negligible compared to the number of parameters of the transformer. The main benefit of 1D RBFs is the encoding of directionality. In terms of how 1-D RBFs are able to sample 3D space using $3m$ points, we would like to clarify that this statement in our previous response was poorly written. Our previous statement intended to convey that to center 1-D RBFs (separately in the x, y, and z direction) to the $m^{3}$ points of a 3D space, you only need to compute $3 m$ distinct 1-D RBFs. All $m^{3}$ RBFs required to map the 3D space can be computed by appropriately multiplying those $3 m $ RBFs. We apologize for the confusion. Thanks again for your feedback and for the insightful discussion, which we believe has substantially improved the quality of our work.
On more careful thinking based on their response, it seems that the 1-D RBF structure authors are using is the same as the standard 3D-RBF. Consider the following argument: The standard 3-D RBF has the form $\exp(-(x-x_0)^2+(y-y_)^2)+(z-z_0)^2)=\exp((x-x_0)^2)\exp((y-y_0)^2)\exp((z-z_0)^2)$ which is the same thing as multiplying 1-D RBFs across the $m^3$ points. So the 3-D RBF should also require the same amount of 3m points as the 1-D RBFs. Regardless, it is a minor point but I wanted to point it out to the authors in case they want to generalize their positional encoding schemes.
Thank you for taking the time to review our manuscript. We are happy you confirm that we have addressed the ethical considerations that are relevant to our work. We are available to provide further clarification, if needed.
To authors
Thank you for the authors' efforts and detailed responses. However, I still have concerns regarding the scale of the dataset. While the authors mention that recent sEEG datasets contain fewer than 12 subjects, this does not sufficiently justify that 23 subjects are adequate. In related fields, datasets often include 50, 100, or even more subjects to ensure robustness and generalizability. I have raised my rating to 4.
I agree that a larger dataset is always preferable. However, for educational purposes only, I must emphasize that collecting a sizable dataset of 50 to 100 subjects is extremely challenging. The first line of treatment for epilepsy typically involves anti-epileptic drugs. Only when these treatments fail after multiple attempts are patients considered for intracranial EEG monitoring. Depending on the size of the epilepsy center, gathering this 50-100 sEEG subjects could take years, if not decades, not to mention the additional difficulty of recruiting them for behavioral studies.
Official Comment by Reviewer pdN2
Thank you for your reminder, and I acknowledge the challenges of data collection from a medical perspective. However, there are still larger, publicly available sEEG datasets. If collecting sEEG data in epilepsy is too difficult, it might be worth considering using other types of brain signals, which could be more suitable for applying machine learning techniques.
Decision
Accept (poster)