UniEEG: Advancing Universal EEG Representation with Electrode-Wise Time-Frequency Pretraining

1.There is a significant issue with the claim of innovation: The electrode-wise preprocessing and mask-based mutual reconstruction methods proposed for constructing the large-scale EEG model are very similar to those in LaBraM[1]. However, the article claims these methods are newly proposed without citing this reference. Notably, this reference is a spotlight paper of ICLR 2024 and is well-known in the field. The lack of corresponding research is unacceptable.
2.Lack of technical depth: For a scientific paper, the absence of any equations throughout the text is problematic. Simple descriptions fail to provide detailed explanations. Although the use of time-frequency features is mentioned, the methods of extracting and unifying features from different datasets are not thoroughly described.
3.Unconvincing experimental results: Only two baseline comparisons are provided, one from 2018 and another from 2015, without comparisons to any recent models in the EEG field, including LaBraM.
4.Formatting issues: The font in Table 1 is significantly reduced, making it difficult to read; the line spacing for headings in Sections 5.3.2 and 5.3.7 is reduced, even overlapping with the text, which does not meet ICLR standards.

[1]Wei-Bang Jiang, Li-Ming Zhao, and Bao-Liang Lu. Large brain model for learning generic representations with tremendous eeg data in bci. arXiv preprint arXiv:2405.18765, 2024.

• Poor Writing quality. The manuscript suffers from numerous writing and formatting issues. For instance, there are two "Introduction" sections on Page 1. Section 4.1.2 contains an incomplete sentence: “most lengths It should be” is missing a period. Similarly, Section 5.3.1 has an incomplete sentence: “To investigate the effect of data domain on.” Additionally, Section 5.3.2 contains “To investigate the representational ability of,” and in Section 5.3.4, “In our” is left hanging. These errors make the paper difficult to follow and detract from its overall quality.
• Insufficient related work. The authors claim in Section 3.2 that "there are no studies validating MSM in EEG signals, which is the main focus of our work." However, several recent studies have already explored the use of masked signal modeling (MSM) for EEG, such as [1], [2], and [3]. These works should be discussed and compared to clarify how UniEEG advances the field.
• Lack of baseline comparison. The abstract and introduction have significant overlap with LaBraM [1], a pioneering foundation model for EEG signals. However, UniEEG does not include a comparison with LaBraM or other commonly used models like BIOT [4]. Instead, the paper only compares UniEEG against its own variations and two out-of-date baselines, which is insufficient to demonstrate its advantages or novelty.
• Limited novelty. The approach of UniEEG is largely a straightforward application of vanilla MAE (Masked Autoencoder) on EEG data. There isn’t a significant methodological innovation beyond adapting a pre-existing model to EEG.
• Missing implementation details. Essential details, such as the training, validation, and test split, are absent. This information is critical for replication and for assessing the reliability of the reported results.

Overall, due to the poor writing, limited novelty, insufficient related work, and lack of robust comparisons, this paper falls short of the standards for a top-tier ML conference like ICLR. However, the ablation studies show some interesting findings. I suggest the authors make substantial revisions and consider submitting the work to a journal instead.

[1] Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI. ICLR 2024.

[2] EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs. KDD 2024.

[3] Neuro-BERT: Rethinking Masked Autoencoding for Self-Supervised Neurological Pretraining. IEEE Journal of Biomedical and Health Informatics 2024.

[4] BIOT: Biosignal Transformer for Cross-data Learning in the Wild. NeuIPS 2023.

1. The general writing quality can be improved a lot. For example, there are two introduction sections, missing spaces in section1, 4.1.1, sentence starting with 'And' in section 5.1.1, grammatical errors in section 5.2, extra spaces before period and missing words in section 5.3.1, 5.3.2, 5.3.4, format error in section 5.3.2.
2. The paper claimed to be 'the first electrode wise time-frequency pretraining model' which unfortunately not the case. Please check the following references: Yang, C., Westover, M., & Sun, J. (2024). Biot: Biosignal transformer for cross-data learning in the wild. Advances in Neural Information Processing Systems, 36. Yi, K., Wang, Y., Ren, K., & Li, D. (2024). Learning topology-agnostic eeg representations with geometry-aware modeling. Advances in Neural Information Processing Systems, 36.
3. The 'resize' operation for segments with random duration is unclear. Is it done through padding or re-sampling? If so, does it mean that the input segments have different sampling rate?
4. Table 1 is too small and unreadable.
5. It is unclear if the SOTA results were obtained through re-implementation and experiments under the same setup or simple report of their performance from their original paper. The s.t.d. of the performance metrices should be provided and significance figures should be reported when making comparison. How were the holdout validation set designed?
6. The data split for evaluation is unclear. Section 5.1.1 reported that there were 18 datasets used in total, 16 were used for pre-training and 12 were used for evaluation. Were the datasets split based on the subject's identity or sessions? This part need more clarification.

The submission presents significant challenges in readability and comprehension. It appears that the authors may lack an understanding of EEG processing issues and propose generalized solutions that lack clarity and relevance.