EEGMamba: Bidirectional State Space Model with Mixture of Experts for EEG Multi-task Classification

Reply to W1-W2

Thank you for your detailed and thoughtful review of our manuscript. We value your comments and suggestions, and we are grateful for the opportunity to address them. In this response, we will provide further clarification and empirical evidence to address the concerns you raised. Here is our detailed response: **W1: The generalizability and flexibility of the current method in handling unforeseen tasks with different channel configurations.** Our model can be extended to new datasets after training. We can describe the specific extension from two different situations as follows: - When the newly added dataset has the same number of tasks, channels and classes as the previous dataset, we can directly use the task index of the original dataset. For example, the two datasets SleepEDF20 and SHHS used in the manuscript experiment can be replaced with each other, that is, if you want to apply the model trained on SleepEDF20 to SHHS, you only need to encode the same task and make a few epochs of fine-tuning. - When the newly added dataset cannot meet the conditions in 1, we need to give it a new task number and pre-set its number of channels and classes, and then a few epochs of fine-tuning. This is broadly similar to what most current foundation models do, except that we need to pre-set the task number and the number of channels. We use the Confused student EEG brainwave data [1] (hereinafter referred to as Confused EEG), which is a completely new task for EEGMamba. We applied the existing weights (i.e., the weights corresponding to Table 2-5 in the manuscript) to the Confused EEG, using 4 different random numbers to obtain a 7:3 training-test set ratio, similar to the approach in [3] (while [2] used five-fold cross-validation). We trained for 10 epochs each time and averaged the results. The test results are shown as follows: | Classification Network | [2], 2019 | [3], 2023 | EEGMamba | |:------------------------:|:---------:|:---------:|:--------:| | Reported Accuracy | 0.7500 | 0.7670 | 0.7825 | [1] Wang H, Li Y, Hu X, et al. Using EEG to Improve Massive Open Online Courses Feedback Interaction[C]//AIED workshops. 2013. [2] Wang H, Wu Z, Xing E P. Removing confounding factors associated weights in deep neural networks improves the prediction accuracy for healthcare applications[C]//Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing. NIH Public Access, 2019, 24: 54. [3] Lim Z Y, Neo Y L. Confused vs Non-Confused Electroencephalography Signal Classification Using Deep Learning Algorithm[C]//2023 IEEE International Conference on Automatic Control and Intelligent Systems (I2CACIS). IEEE, 2023: 195-200. **W2: The motivation for employing Mamba, especially Bidirectional Mamba, in this work is not sufficiently clear or logically aligned with this objective.** In this work, we use the bidirectional Mamba for feature extraction from EEG signals, not only to reduce computational complexity but also to enhance the model’s overall ability to capture the characteristics of EEG signals. **Let us be crystal clear: adopting bidirectional modeling does not affect the real-time detection capabilities of EEG signals.** In reality, whether it’s single-directional or bidirectional modeling, what we’re processing are already sampled EEG signal segments—not like in natural language tasks, where you need to predict the next word. Therefore, bidirectional modeling has no negative impact on real-time performance. As we know, Transformer-based models typically suffer from quadratic computational complexity, which leads to significant performance bottlenecks in practical applications. Take medical diagnostic scenarios, for example, such as sleep disorder detection or Alzheimer’s diagnosis, where you need to process long durations (e.g., several minutes) and a large number of channels (e.g., 62 channels) of EEG data. In such cases, Transformer-based models consume an enormous amount of GPU memory, making it nearly impossible to meet the real-time processing requirements. On the other hand, the Mamba model, with its linear computational complexity, consumes much less memory when handling long sequences, making it far better suited for high-load tasks like these. This is the primary reason we chose the Mamba model. The reason for choosing bidirectional Mamba is that the original Mamba model was designed for language generation tasks, which typically use single-directional modeling. But the nature of EEG signals makes single-directional modeling unsuitable—it could lead to the loss of earlier information during the scanning process. To more comprehensively and accurately capture the temporal features of EEG signals, we employed a bidirectional Mamba structure. Our ablation experiments also well demonstrated the performance difference between single-directional modeling and bidirectional modeling in EEG classification.

Reply to W3-W6

**W3: In Section 4.1, the authors present the performance of the single-task EEGMamba and other transformer-based models concerning memory usage and inference speed as the sequence length increases…** From Figure 4, it is clear that, for other Transformer-based models, the blue line representing the Single-task EEGMamba shows a distinct advantage. The only model that can somewhat match Single-task EEGMamba when processing single-channel data is HCANN, but in multi-channel scenarios, its performance rapidly declines. Meanwhile, all the other models consistently perform worse than Single-task EEGMamba. As for the signal length, the actual sequence length for each signal can be obtained from the rate × duration in Table 1. We will consider replacing "duration" with "sequence length" to make it easier for the readers to understand. Clearly, EEGNet’s overall performance on the dataset we’re using is far from comparable to EEGMamba, especially when it comes to tasks like sleep stage detection, which involves longer signals. Additionally, EEGNet requires dedicated training for each dataset, which is quite a tedious process. **W4: In Section 4.2, the authors present the performance of EEGMamba in multi-task classification…** EEGMamba has been evaluated on eight publicly available EEG datasets across four different tasks, demonstrating superior performance in seizure detection, emotion recognition, sleep stage classification, and motor imagery. Our model ranks among the top three on seven datasets and achieves the best performance on four datasets, outperforming existing state-of-the-art (SOTA) models across multiple datasets. In addition, EEGMamba is an end-to-end system that does not require separate pre-training and fine-tuning stages, offering stronger generalization ability than pre-trained models. Unlike other classification networks that require multiple training sessions with manual adjustments to data length, channel count, and class numbers, EEGMamba only needs to be trained once to achieve these results. **W5: Regarding Figure 5, I observe that, apart from the sleep stage task…** When it comes to the activation probabilities of experts, it’s not only important to observe the differences in activation probabilities between experts for a given task, but even more crucial to focus on the tendency of experts being activated across different tasks. For instance, in the task of seizure detection, experts 5 and 6 appear to be more easily activated, while for tasks like emotion recognition and motor imagery, experts 2 and 4 are more likely to be chosen, respectively. This suggests that different tasks tend to favor different experts. In other words, each task seems to have its own preference for specific experts, which is an interesting characteristic that could be leveraged to improve task-specific performance. **W6: Figure 6 presents the ablation study results for the various modules of EEGMamba…** We sincerely apologize, but due to the particularly poor performance of certain variants (such as the unidirectional Mamba), it’s been difficult to further adjust the y-axis range in a way that would make the differences more visually distinct. Additionally, for some datasets and tasks, the inherent characteristics of the dataset (for example, the extreme class imbalance in the epilepsy dataset) make it quite challenging to observe significant differences when evaluating using accuracy. We recommend that you enlarge the image up to 400% so that you’ll be able to more clearly see the differences between the various variants.

Reply to Q1-Q3

**Q1: Could you please specify how the EEG tasks are encoded into task tokens?** **A1:** We sincerely apologize for any confusion caused by the unclear presentation. The task token is pre-assigned to each task before model training begins. When the EEG feature tokens pass through the task-aware gate, both the task token and the EEG feature tokens are both involved in the computation of the task-aware gate. **Q2: I noticed that the DEAP dataset was utilized in this study, but only data from four electrodes were selected… What is the reasoning for selecting such unusually long data lengths?** **A2:** As described in Appendix D5, we selected four channels for classification based on the approach outlined in [1], where grid search was used to achieve similar results. More detailed information on this procedure can be found directly in [1]. Thank you for pointing out that our manuscript omitted a description of the emotional evaluation dimension used. Specifically, we employed valence as the indicator for binary classification, and we will include this information in the revised version of the paper. Regarding the selection of duration, we found that previous literature does not provide direct evidence supporting the choice of 1s, 2s, or 4s as the optimal segment lengths [2][3]. It is likely that these choices were made due to limitations in Transformer-based models, which struggled to classify longer EEG sequences due to constrained computational resources. [1] Khateeb M, Anwar S M, Alnowami M. Multi-domain feature fusion for emotion classification using DEAP dataset[J]. IEEE Access, 2021, 9: 12134-12142. [2] Song Y, Zheng Q, Liu B, et al. EEG conformer: Convolutional transformer for EEG decoding and visualization[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2022, 31: 710-719. [3] Jiménez-Guarneros M, Fuentes-Pineda G. Cross-subject EEG-based emotion recognition via semi-supervised multi-source joint distribution adaptation[J]. IEEE Transactions on Instrumentation and Measurement, 2023. **Q3: This work employs five-fold cross-validation for data partitioning, which does not appear to be a commonly used EEG dataset partitioning method. What is the rationale or basis for this choice?** **A3:** We apologize for any confusion, but as far as we know, subject-wise cross-validation is a widely used method for partitioning EEG datasets, with common approaches including five-fold, ten-fold, or even twenty-fold cross-validation. References supporting this include [1][2][3][4][5]. We are curious to know the rationale behind your statement that this is an uncommon form of data partitioning. [1] Eldele, E., Chen, Z., Liu, C., Wu, M., Kwoh, C. K., Li, X., & Guan, C. (2021). An attention-based deep learning approach for sleep stage classification with single-channel EEG. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29, 809-818. [2] Yuan Z, Zhang D, Chen J, et al. Brant-2: Foundation Model for Brain Signals[J]. arXiv preprint arXiv:2402.10251, 2024. [3] Ji Y, Li F, Fu B, et al. A novel hybrid decoding neural network for EEG signal representation[J]. Pattern Recognition, 2024, 155: 110726. [4] Lawhern V J, Solon A J, Waytowich N R, et al. EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces[J]. Journal of neural engineering, 2018, 15(5): 056013. [5] Zhang Z, Zhong S, Liu Y. TorchEEGEMO: A deep learning toolbox towards EEG-based emotion recognition[J]. Expert Systems with Applications, 2024, 249: 123550.

A reminder to reviewerSRQL

Dear Reviewer SRQL, We would like to gently remind you that we have submitted our response to the concerns you raised. We hope these further explanations help to resolve any outstanding issues. We sincerely hope that you might consider revising your score based on our response. If you have any questions, we would be more than happy to discuss with you. Thank you again for the time and effort you put into reviewing our manuscript. Best regards, The authors

Reply to Q1

Thank you for your insightful comments. We appreciate the opportunity to provide further clarification and evidence to support our claim. We have noticed that the weakness you mentioned has a corresponding relationship with the question, so we will answer the question directly. Here is our detailed response: **Q1. Introduction: (line 49-50):** **Q1.a: For EEG classification, what defines a "long" signal in terms of sample size? At what point does a short receptive field cause CNN performance to degrade?** **A1.a:** In the context of our manuscript, "long" refers to EEG signals that exceed the typical short segments (e.g., several seconds) that CNNs are traditionally designed to handle well. Our claim is based on the observation that as the sequence length increases, the performance of CNN-based models, which lack a mechanism to explicitly model long-range dependencies, can degrade. For example, EEG signals for sleep monitoring must have a standard length of 30s, which tends to result in a larger length, and we also observed that CNN-based EEGNet did not perform very well on the sleep stage classification task. To clarify, the degradation in performance is not a hard threshold but a gradual effect that becomes more pronounced with longer sequences. The exact point at which performance degrades can vary based on the specific architecture, the complexity of the task, and the characteristics of the EEG signals. **Q1.b: Is global sequence modelling truly necessary for long-term EEG signals? From a neuroscience perspective, how much does brain activity from, say, 10 seconds ago, affect the current state? Which types of tasks specifically require such long-term modelling?** **A1.b:** Global sequence modeling is not necessary for every EEG task, but it is essential for tasks where the temporal dynamics of brain activity play a key role in accurate classification. From a neuroscience perspective, while the effects of brain activity may diminish over time, certain types of brain activity (such as those associated with memory recall or seizures) may have lasting effects. On the other hand, long-term patterns in some states are crucial for understanding brain states, such as sleep stage classification. In sleep stage classification, longer EEG segments (such as 30 seconds) are required due to the cyclical nature of the sleep cycle and the evolving pattern of brain activity. During sleep, especially in the deep sleep stages, EEG features tend to be slow and persistent. Shorter clips may not be enough to capture a steady pattern of these activities, while 30-second segments often accurately reflect the brain’s electrical activity at each sleep stage. Therefore, for tasks such as sleep stage classification, global sequence modeling, which captures both short-term and long-term dependencies, is essential for accurate classification.

Reply to Q2

**Q2. section 2.2 ST-Adaptive Module:** **Q2.a: Scalability to New Datasets.** **A2.a:** Our model can be extended to new datasets after training. We can describe the specific extension from two different situations as follows: - When the newly added dataset has the same number of tasks, channels and classes as the previous dataset, we can directly use the task index of the original dataset. For example, the two datasets SleepEDF20 and SHHS used in the manuscript experiment can be replaced with each other, that is, if you want to apply the model trained on SleepEDF20 to SHHS, you only need to encode the same task and make a few epochs of fine-tuning. - When the newly added dataset cannot meet the conditions in 1, we need to give it a new task number and pre-set its number of channels and classes, and then a few epochs of fine-tuning. This is broadly similar to what most current foundation models do, except that we need to pre-set the task number and the number of channels. We use the Confused student EEG brainwave data [1] (hereinafter referred to as Confused EEG), which is a completely new task for EEGMamba. We applied the existing weights (i.e., the weights corresponding to Table 2-5 in the manuscript) to the Confused EEG, using 4 different random numbers to obtain a 7:3 training-test set ratio, similar to the approach in [3] (while [2] used five-fold cross-validation). We trained for 10 epochs each time and averaged the results. The test results are shown as follows: | Classification Network | [2], 2019 | [3], 2023 | EEGMamba | |:------------------------:|:---------:|:---------:|:--------:| | Reported Accuracy | 0.7500 | 0.7670 | 0.7825 | [1] Wang H, Li Y, Hu X, et al. Using EEG to Improve Massive Open Online Courses Feedback Interaction[C]//AIED workshops. 2013. [2] Wang H, Wu Z, Xing E P. Removing confounding factors associated weights in deep neural networks improves the prediction accuracy for healthcare applications[C]//Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing. NIH Public Access, 2019, 24: 54. [3] Lim Z Y, Neo Y L. Confused vs Non-Confused Electroencephalography Signal Classification Using Deep Learning Algorithm[C]//2023 IEEE International Conference on Automatic Control and Intelligent Systems (I2CACIS). IEEE, 2023: 195-200. In addition to this, we need to state that the starting point of our model design is an end-to-end multi-task classification model. **Q2.b: Tokenize Layer.** **A2.b:** The choice of kernel sizes in the tokenize layer is mainly based on the methods of existing papers and our own experimental results. We refer to the MRCNN module in AttnSleep [1], which has two branches of different kernel sizes in order to enable the model to learn features at multiple scales simultaneously. The small kernel captures local, high-frequency details, while the wide kernel captures more global, low-frequency patterns. This dual-path approach is common in many EEG signal processing architectures [2][3]. Therefore, we adopt the same structural design and selected the final convolution kernel size after a series of ablation studies. In these studies, we tested a variety of kernel sizes and found that the sizes now employed in the paper provide a good balance between capturing fine-grained and broader spatial features in EEG signals. [1] Eldele E, Chen Z, Liu C, et al. An attention-based deep learning approach for sleep stage classification with single-channel EEG[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2021, 29: 809-818. [2] Zhu H, Zhou W, Fu C, et al. Masksleepnet: A cross-modality adaptation neural network for heterogeneous signals processing in sleep staging[J]. IEEE Journal of Biomedical and Health Informatics, 2023, 27(5): 2353-2364. [3] Zhu H, Wang L, Shen N, et al. MS-HNN: Multi-scale hierarchical neural network with squeeze and excitation block for neonatal sleep staging using a single-channel EEG[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2023, 31: 2195-2204.

Reply to Q3

**Q3. section 3.2 Data Division:** **Q3.a: Subject Variability Impact.** **A3.a:** You are correct that subject variability is a significant challenge in BCI tasks, particularly in motor imagery. The lower performance on the BCI-IV-2a and SEED datasets, as compared to seizure detection and sleep stage classification, is indeed due to the high subject variability in these tasks. While the performance on these specific datasets may not reach the levels seen in subject-specific models, it is important to note that EEGMamba is trained to to work across various subjects without the need for individual calibration. We believe this is a valuable property for clinical applications where subject-specific models may not be feasible. **Q3.b: Performance Discrepancy with Benchmark Models.** **A3.b:** The performance drop of the EEG Conformer and other benchmark models when evaluated using a subject-split approach highlights the sensitivity of these models to subject variability and the challenges of subject transfer. EEGMamba, with its multi-task learning approach, is designed to be more robust to such variability. We supplemented EEGMamba’s performance on BCI-IV-2a in the case of subject-specific training, and the standard deviation in the experiment was derived from five different random numbers. It should be noted that the results in EEGConformer were obtained through data augmentation techniques of segmentation and reconstruction, while all the experimental results in EEGMamba’s paper did not carry out similar data augmentation techniques. And in order to maintain unity with the original manuscript, we did not do this in our experiment. | methods | s01 | s02 | s03 | s04 | s05 | s06 | s07 | s08 | s09 | average | |:--------:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-------:| | EEGMamba | 80.35 | 64.18 | 85.38 | 74.89 | 75.53 | 61.35 | 87.37 | 84.11 | 74.54 | 76.41 | While we cannot guarantee that EEGMamba would match subject-specific models like EEG Conformer when evaluated on a single subject, our model’s strength lies in its ability to generalize across subjects. We acknowledge the need for further investigation into how EEGMamba performs in a subject-specific setting and will include comparative analyses in future work. **Q3.c: Use of Separate Test Sets.** **A3.c:** We apologize for the lack of clarity regarding the use of separate test sets. In our experiments, we did not use the official training set and test set provided by BCI-IV-2a for final model training and evaluation. The reason for this is because we want the main results to have a common experimental setting and show cross-subject generalization of the model, so we use a subject-split setting. As mentioned in A3.b, we supplemented the experimental results obtained using the official training set and test set of BCI-IV-2a. We will clarify this in the revised manuscript to ensure transparency and to facilitate fair comparisons with previous work.

Reply to Q4-Q6

**Q4: section 4.1 Single-Task EEGMamba Performance Comparison.** **A4:** As mentioned in A1.a, the term "long" in the context of sequence length for Transformer does not refer to a strict threshold, but rather to a gradual point at which the quadratic computational complexity of transformers becomes a practical limitation. This threshold can vary depending on the specific model architecture and the available computational resources. It is important to note that the length of the sequence used for model processing is not directly equal to the duration of the signal, but is also related to the sampling frequency. For example, for a sleep-state EEG with a duration of 30 seconds and a sampling rate of 200Hz, its length is 6,000 data points. For some Transformer-based models, this has led to the need for larger memory-usage and a sharp drop in reasoning speed. **Q5: section 4.2 EEGMamba for EEG Multi-Task Classification.** **A5:** Multi-task training significantly improves the model’s ability to generalize across different datasets and tasks. While single-task models may excel on specific datasets, they often fail to achieve this across multiple tasks or datasets. By leveraging shared representations, EEGMamba can transfer knowledge from one task to improve performance on others, similar to how some foundation models are pre-trained to learn a common representation of EEG data and then fine-tuned for specific tasks. However, EEGMamba goes beyond this by integrating the entire process into an end-to-end system, allowing it to learn task-specific features while still benefiting from shared knowledge. Regarding the current experimental results, we acknowledge that they are based on the mean and standard deviation from five-fold cross-validation, which involves a limited number of experiments. As a result, this may affect the statistical significance of the P-value. Like most AI studies [1][2][3], we did not report the statistical differences in the results. Nevertheless, we believe that while multi-task training may not always lead to improved performance on each individual dataset, it provides a more robust framework for handling a variety of tasks. [1] Zhang D, Yuan Z, Chen J, et al. Brant-X: A Unified Physiological Signal Alignment Framework[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024: 4155-4166. [2] Yang C, Westover M, Sun J. Biot: Biosignal transformer for cross-data learning in the wild[J]. Advances in Neural Information Processing Systems, 2024, 36. [3] Jiang W, Zhao L, Lu B. Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI[C]//The Twelfth International Conference on Learning Representations. **Q6: Lines 454-456** **Q6.a: Cross-Session Issues:** **A6.a:** We apologize for any confusion caused by our lack of clarity. EEGMamba is capable of learning features from multiple datasets within a single training session, but this applies only to the datasets included in the training set, rather than to completely unseen datasets in a zero-shot setting. Therefore, for new, unseen datasets, fine-tuning as described in A2.a is still necessary. **Q6.b: Generalization to New Datasets:** **A6.b:** We described the method of using EEGMamba for new data in A2.a. On the one hand, if SHHS is not applied to training, it can be assigned the same task label as SleepEDF20 because its number of channels and classes are equal to SleepEDF20. For BCI-IV-2B, which has fewer channels, we need to give it a new task number and fine-tune it a few epochs, because the trained model includes the ability to extract motion imagination features, which is still fundamentally different from training a model from scratch. **Q6.c: Domain-Specific Advantages:** The benefits of multi-task training are not strictly domain-specific but are related to the similarity and diversity of the tasks. For tasks with overlapping features, such as different types of brain-computer interface (BCI) applications, multi-task training can be advantageous. **Q6.d: Practicality of Multi-Task Training:** The practicality of multi-task training with EEGMamba is that it provides a flexible framework that can be adapted to various research settings. For researchers focusing on a single task, EEGMamba can still be applied by treating it as a multi-task model with a single task. This approach can leverage the model’s ability to learn from any available related data, potentially improving performance and generalization. In scenarios where multiple tasks are relevant, such as simultaneous classification of motor imagery and emotion, EEGMamba’s multi-task training provides a more efficient and integrated solution.

Response to weakness and questions

Thanks. It seems there has been a misunderstanding on the definition of a multitask classification model, so perhaps a brief clarification is in order. **As we clearly stated in the paper, our model is designed to handle classification tasks across multiple datasets simultaneously, without the need for the fine-tuning training on each dataset separately**—unlike the base models cited by reviewer wapT. Therefore, the references provided by reviewer wapT [1, 2, 5, 6] clearly do not represent multitask models, as they require dedicated fine-tuning for each specific dataset. **W1: The author claims that EEGmamba is the first universal EEG classification network to effectively implement multi-task learning for EEG applications. However, several established methods are available that can be applied to multi-task EEG learning, as cited in [1][2][3].** As we have stated in a previous clarification, [1, 2] are not multitask models. [3] is a multitask model. However, since it was published recently (within a month of submission deadline) and has not yet undergone peer review, we will not consider it for comparison at this time. Moreover, we have noticed that the paper submitted to this conference, *NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals*, bears a striking resemblance to [3]. And we have observed that several reviewers have already expressed concerns about the performance of [3]: "4vcv: **NeuroLM’s performance is currently lower compared to LaBraM on most tasks**" and "2GzD: **While NeuroLM performs well in multi-task learning, there are still gaps in its performance on some tasks compared to state-of-the-art models specifically designed and optimized for a single task (e.g., LaBraM)**" We fully agree with these assessments, and we have added a comparison with LaBraM. However, **let it be clear: we are not comparing LaBraM as a multitask model similar to ours**, but rather as another baseline comparable to BIOT. [3] can also support us in doing so by the fact that Tables 2-4 in [3] give LaBraM a ‘×’ for multitask performance. Therefore, we sincerely hope that reviewer wapT could take the time to carefully read the relevant literature, and consider change the review points. | Classification Network | Multitask Model | Epilepsy detection | | Sleep stages classification | | Emotion recognition | | Motor imagery | | |:------------------------:|:---------------:|:--------------------:|:-------------:|:-----------------------------:|:-------------:|:---------------------:|:-------------:|:---------------:|:-------------:| | | | Siena | CHB-MIT | SleepEDF-20 | SHHS | DEAP | SEED | Shu | BCI-IV-2a | | LaBraM | × | 0.9886±0.0043 | 0.9742±0.0099 | 0.7503±0.0388 | 0.7785±0.0243 | 0.5822±0.0321 | OOM | 0.5368±0.0312 | 0.2879±0.0160 | | Single-task EEGMamba | × | 0.9897±0.0053 | 0.9817±0.0036 | 0.8387±0.0399 | 0.8441±0.0163 | 0.5985±0.0247 | 0.5779±0.0584 | 0.6169±0.0467 | 0.4596±0.0547 | | EEGMamba | √ | 0.9897±0.0038 | 0.9789±0.0132 | 0.8486±0.0276 | 0.8478±0.0177 | 0.5994±0.0134 | 0.5646±0.0366 | 0.6207±0.0505 | 0.4231±0.0522 | **W2: Inappropriate Comparisons: The choice of baselines for comparison lacks relevance. … To evaluate cross-task capabilities, comparisons should involve multi-task learning methods like those in [1][2][3][5][6].** As we have pointed out in a previous clarification, [1, 2, 5, 6] are not multitask models. Among them, [5, 6] are both specifically designed for intracranial EEG signals and only uses SEEG data for pretraining. EEG and SEEG are two distinct modalities, and to my knowledge, there is no literature that suggests the differences between them can be simply ignored. If reviewer wapT can provide such references, we might consider [5, 6] as baselines. Additionally, we noticed that the papers cited by reviewer wapT, [1, 3], also do not compare with the earlier works [5, 6], which also supports our view. **W3: Efficiency Evaluation: The author should provide quantitative evidence demonstrating the proposed method’s efficiency advantages over previous approaches.** On the one hand, Table 2-5 in our manuscript has used quantitative indicators ACC, AUROC, F1 to give the experimental results of EEGMamba compared with previous methods, proving the performance advantage of EEGMamba. On the other hand, the computational efficiency of the Mamba-based model is shown in Figure 4. We are not sure what "quantitative evidence" means, perhaps you could be more specific. The references order in our response is the same as that of the reviewer wapT.

Response to weakness and questions

Thank you for your valuable feedback. We greatly appreciate the time and effort you have dedicated to reviewing our work. The following are our detailed responses to your weaknesses and questions. **W1: In Section 4.1, the discussion on single-channel and multi-channel models only compares memory usage and inference speed, without evaluating the impact of multi-channel models on performance metrics.** Thanks. A comparison of our performance metrics results is shown in Table 2-5, where the number of channels for each dataset is shown in Table 1, with the number of channels ranging from 1 to 62. And we can obtain the EEG signal length from each data set according to the rate×duration in Table 1. At present, for single-task EEGMamba, the maximum length is 128×60=7680 from the DEAP dataset, which we consider to be a very large length. When training with EEGMamba, our operation to unify the sampling frequency to 200Hz resulted in more datasets of larger length, such as the length of DEAP dataset is 200×60=12000, and the length of SHHS dataset is 200×30=6000. EEGMamba has the best performance among all models in this case. **W2: The t-SNE visualization lacks layer-by-layer analysis of the model’s influence on clustering results, which does not adequately demonstrate the feature extraction capability of each layer.** T-SNE is a dimensionality reduction technique that is useful for visualizing high-dimensional data in two or three dimensions, focusing on the relative distances between data points to reflect similarities or dissimilarities. In our study, we use t-SNE to qualitatively assess EEGMamba’s ability to extract discriminative features, showing clear separation between classes in the plot. This clustering indicates that the model effectively distinguishes between different data distributions, supporting our claims about its feature extraction capabilities. However, it is crucial to note that the coordinates in a t-SNE plot lack inherent meaning, emphasizing relative positions rather than absolute values. While layer-by-layer t-SNE visualizations could offer additional insights, they are less informative due to the lack of inherent meaning in the coordinates. Instead, we focus on overall model performance, using metrics like classification accuracy, AUC-ROC, and F1 scores to provide a more direct and quantitative assessment. **Q1: In Section 4.2’s experimental comparison, was the model trained using all the datasets at once? Could there be interactions between the datasets? Would training the model on each dataset separately improve the performance metrics? Have you conducted any experiments on this? I’m quite interested.** **A1:** You are correct in noting that our model uses all the datasets in a single training process, which allows for interactions between them. And in the Single-task EEGMamba experiment, we trained the model separately on different datasets, and the corresponding experimental results are presented in Tables 2-5. We sincerely hope this helps solve your doubts. Interactions between datasets in multi-task learning can have both positive and negative effects. On the positive side, the model benefits from shared knowledge across tasks. For instance, certain brain activity patterns, such as oscillatory rhythms or temporal dynamics, may be consistent across different tasks. However, conflicting features or task-specific noise can interfere with the model’s ability to learn meaningful representations for each individual task. To address these challenges, we implemented several strategies in EEGMamba: **Task-Aware Modules:** A Mixture of Experts (MoE) model activates different experts based on the task, allowing the model to learn task-specific features while leveraging shared knowledge. **Universal Expert:** Another key component of EEGMamba is the universal expert, which is designed to capture the cross-task features that are universally relevant, preventing overfitting to task-specific noise and enables better generalization across datasets. In summary, while interactions between datasets can lead to both positive and negative effects in multi-task learning, our model incorporates strategies such as task-aware modules and the universal expert to better exert positive effects and eliminate negative effects, improving EEGMamba’s performance across different EEG tasks.

Reply to Q1-Q2

Thank you for your detailed and thoughtful review of our manuscript. In this response, we will provide further clarification and empirical evidence to address the concerns you raised. Since we have noticed that the weakness you mentioned has a corresponding relationship with the question, we will answer the question directly. Here is our detailed response: **Q1: Could you elaborate on the choice of baseline models? How does EEGMamba’s performance compare with multi-task EEG models specifically designed for generalization across diverse tasks?** **A1:** We choose EEGNet, AttnSleep, EEGConformer and HCANN as single-task classification models for comparison. BIOT serves as a representative model that is pre-trained and then fine-tuned on a single dataset. And as you suggested, we added a performance comparison with LaBraM, which is the same type model as BIOT. | Classification Network | Multitask Model | Epilepsy | detection | Sleep stage | classification | Emotion | recognition | Motor | imagery | |:------------------------:|:---------------:|:--------------------:|:-------------:|:-----------------------------:|:-------------:|:---------------------:|:-------------:|:---------------:|:-------------:| | | | Siena | CHB-MIT | SleepEDF-20 | SHHS | DEAP | SEED | Shu | BCI-IV-2a | | LaBraM | × | 0.9886±0.0043 | 0.9742±0.0099 | 0.7503±0.0388 | 0.7785±0.0243 | 0.5822±0.0321 | OOM | 0.5368±0.0312 | 0.2879±0.0160 | | Single-task EEGMamba | × | 0.9897±0.0053 | 0.9817±0.0036 | 0.8387±0.0399 | 0.8441±0.0163 | 0.5985±0.0247 | 0.5779±0.0584 | 0.6169±0.0467 | 0.4596±0.0547 | | EEGMamba | √ | 0.9897±0.0038 | 0.9789±0.0132 | 0.8486±0.0276 | 0.8478±0.0177 | 0.5994±0.0134 | 0.5646±0.0366 | 0.6207±0.0505 | 0.4231±0.0522 | The difference between EEGMamba and these models is that: **1. End-to-End Training:** EEGMamba is trained end-to-end, allowing all components of the model to learn jointly, which makes EEGMamba highly efficient, as it avoids the need for multiple rounds of fine-tuning for each task. This is in contrast to models that train a shared backbone separately and then attach task-specific heads, which can be time-consuming and less efficient, especially when working with a large number of diverse tasks. **2. Multi-Task Learning Capability:** EEGMamba is designed to handle multiple tasks and datasets simultaneously within a single training session. Unlike existing foundation models, which can only process one dataset at a time during fine-tuning, EEGMamba is capable of delivering results from multiple datasets simultaneously in a single session. In summary, EEGMamba offers a more integrated and sophisticated approach to multi-task learning for EEG applications. We hope this clarification highlights the unique advantages of EEGMamba and addresses your query effectively. **Q2: How does the ST-Adaptive module capture interactions between channels and temporal features, which are critical in EEG data? Have you considered modeling these dependencies more explicitly?** **A2:** The ST-Adaptive module sequentially extracts the spatial features and temporal features of the EEG signals, without direct interactions. First, it uses a spatially adaptive convolution method to normalize EEG data to a fixed number of channels, enabling the model to capture channel interactions despite spatial diversity across datasets and tasks. Then the extracted channel features are input into the dual path structure in the tokenizer, which includes both small and wide kernel convolutions, allows the model to capture local and global temporal features. The resulting token naturally contains channel features and time features, and does not require additional spatio-temporal interaction.

Reply to Q3-Q5

**Q3: How would EEGMamba perform on out-of-distribution tasks or tasks not seen during training? Have you tested its ability to generalize to entirely new EEG task types?** **A3:** Thank you for the good suggestion. We use the Confused student EEG brainwave data [1] (hereinafter referred to as Confused EEG), which is a completely new task for EEGMamba. We applied the existing weights (i.e., the weights corresponding to Table 2-5 in the manuscript) to the Confused EEG, using 4 different random numbers to obtain a 7:3 training-test set ratio, similar to the approach in [3] (while [2] used five-fold cross-validation). We trained for 10 epochs each time and averaged the results. The test results are shown as follows: | Classification Network | [2], 2019 | [3], 2023 | EEGMamba | |:------------------------:|:---------:|:---------:|:--------:| | Reported Accuracy | 0.7500 | 0.7670 | 0.7825 | [1] Wang H, Li Y, Hu X, et al. Using EEG to Improve Massive Open Online Courses Feedback Interaction[C]//AIED workshops. 2013. [2] Wang H, Wu Z, Xing E P. Removing confounding factors associated weights in deep neural networks improves the prediction accuracy for healthcare applications[C]//Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing. NIH Public Access, 2019, 24: 54. [3] Lim Z Y, Neo Y L. Confused vs Non-Confused Electroencephalography Signal Classification Using Deep Learning Algorithm[C]//2023 IEEE International Conference on Automatic Control and Intelligent Systems (I2CACIS). IEEE, 2023: 195-200. **Q4: Could you clarify how this work builds on or differs from previous applications of SSMs or Mamba models in EEG classification? Adding this context could help readers better understand the specific contributions of EEGMamba.** **A4:** We searched for relevant keywords in Google Scholar and reviewed the literature published before the paper submission deadline. Most of the studies applying Mamba to EEG are still in preprint form and have not yet undergone peer review. Therefore, we provide a brief summary here to give readers an overview of the application of SSMs or Mamba models in EEG classification. [1] is the first application of Mamba to EEG used a single-directional Mamba model, but the article provides only a brief introduction to the methodology and experiments. [2] proposed a self-supervised learning (SSL) framework for sleep stage classification, utilizing a Mamba-based temporal context module to capture relationships among different EEG epochs, although the main component is still a Transformer-based model. [3] used Brain Timeseries Mamba and Brain Network Mamba module to encode the spatiotemporal features and long-term dependencies of EEG signals, respectively, to achieve efficient EEG classification. [1] Panchavati S, Arnold C, Speier W. Mentality: A Mamba-based Approach towards Foundation Models for EEG [J]. [2] Lee C H, Kim H, Han H, et al. NeuroNet: A Novel Hybrid Self-Supervised Learning Framework for Sleep Stage Classification Using Single-Channel EEG[J]. arXiv preprint arXiv:2404.17585, 2024. [3] Behrouz A, Hashemi F. Brain-mamba: Encoding brain activity via selective state space models[C]//Conference on Health, Inference, and Learning. PMLR, 2024: 233-250. **Q5: Can you evaluate this on longer contexts to better get a sense for the necessity and usefulness for Mamba?** **A5:** Thank you for your suggestion. We can obtain the EEG signal length from each data set according to the rate×duration in Table 1. At present, for single-task EEGMamba, the maximum length is 128×60=7680 from the DEAP dataset, which we consider to be a very large length. When training with EEGMamba, our operation to unify the sampling frequency to 200Hz resulted in more datasets of larger length, such as the length of DEAP dataset is 200×60=12000, and the length of SHHS dataset is 200×30=6000. EEGMamba has the best performance among all models in this case.