EAD: An EEG Adapter for Automated Classification

Aditya Nigam; Arnav Bhavsar; Jyoti Nigam; Medicherla Vamsi Krishna; Pushapdeep Singh

REVIEW 3 major objections 5 minor 30 references

EAD: An EEG Adapter for Automated Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An EEG adapter distills any montage into one foundation model, reaching 99.33 percent on EEG-ImageNet.

desk verdict Useful adapter idea with plausible main results, but the zero-shot generalization claim does not survive contact with Table 3. read the letter →

arxiv 2505.23107 v1 pith:I52M2EEB submitted 2025-05-29 cs.CV cs.AI

Pushapdeep Singh , Jyoti Nigam , Medicherla Vamsi Krishna , Arnav Bhavsar , Aditya Nigam This is my paper

classification cs.CVcs.AI

keywords EEGclassificationfoundationmodelLaBraMtemporalconvolutionadaptervisualbraindecodingresting-statesubject-independentzero-shot

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a single, device-agnostic pipeline can classify EEG from very different experiments without task-specific preprocessing or manual channel-location mapping. The key move is a temporal-convolution adapter, EAD, that automatically distills an arbitrary multichannel EEG sample into the fixed input space of LaBraM, a pretrained EEG foundation model. On EEG-ImageNet, the approach reaches 99.33% classification accuracy, surpassing the published state of the art; on the resting-state BrainLat dataset, it reaches 92.31% subject-level accuracy in a subject-independent setup. The paper also reports 98.21% accuracy on six classes excluded from training, in a zero-shot visual-decoding test. If the result holds, one trained backbone plus a light adapter could replace the usual practice of building a new network for every EEG device and task.

What carries the argument

The load-bearing object is the EEG Adapter (EAD): a small CNN using temporal convolutions whose only job is to reshape a raw multichannel EEG sample of arbitrary channel count and length into the fixed input dimensions of the Base Foundation Model (BFM), here LaBraM. LaBraM is a pretrained EEG foundation model that learns generic representations via a VQ-VAE codebook and a masked transformer; EAD sits in front of it and learns which temporal signal information to keep, while the BFM is fine-tuned to learn temporal and spatial dependencies. The paper also implements baseline channel alignment strategies, manual electrode selection and manual mixing of surrounding channels, to show that EAD's automatic distillation gives the best accuracy.

What would settle it

Train EAD on EEG-ImageNet's 128-channel montage and then, with no adapter retraining, apply it directly to BrainLat's resting-state montage; if accuracy falls to near chance, the claimed device-agnostic universal pipeline fails. A second test: record the same visual task on a low-density device, for example 8 or 16 channels, and check whether a single EAD reaches the reported accuracy without per-dataset architecture changes.

Watch

Extended reading notes

Core claim

The central claim is that a learned temporal-convolution adapter placed in front of a pretrained EEG foundation model is enough to make EEG classification montage-agnostic. EAD takes an $E \times T$ EEG signal and maps it to the $I_{Ch}$ by $I_T$ input shape LaBraM expects, using categorical cross-entropy loss; the adapter and backbone are fine-tuned end to end. The paper shows that manual anatomical channel matching or mixing becomes unnecessary, and that the automatically distilled representation separates classes better than manual channel selection in t-SNE plots. With this setup, EAD outperforms existing EEG-classification models on visual stimulus decoding, reporting 99.33% on EEG-ImageNet, and presents subject-independent resting-state results on BrainLat at 92.31% subject-level accuracy.

Load-bearing premise

The load-bearing premise is that LaBraM's pretrained feature space is universal enough that a small CNN with no electrode-location input can map any montage into it without losing the signal that discriminates classes.

Editorial extensions

If this is right

EEG classification can move from per-task, per-device pipelines to one foundation backbone plus a light adapter, reducing preprocessing and manual montage matching.
Subject-independent resting-state diagnosis, distinguishing Alzheimer's disease, behavioral variant frontotemporal dementia, and healthy controls, is feasible at 92.31% subject-level accuracy without using subject identity in training.
Zero-shot decoding works for visual classes never seen in training: EAD features plus an SVM reach 98.21% on six held-out EEG-ImageNet classes.
The same 50 Hz notch and 0.1-75 Hz bandpass preprocessing serves both stimulus-based and resting-state datasets, suggesting a common preprocessing front end.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: a direct test the paper leaves implicit is to train EAD on one montage and apply it to another montage with no adapter retraining; if accuracy holds, EAD is genuinely montage-invariant, and if it collapses, the universality claim is limited to per-dataset retraining.
Editorial inference: the success of automatic distillation suggests that explicit electrode-location priors may be unnecessary once a foundation model supplies a rich feature space, which could motivate channel-agnostic adapters for other biosignals such as MEG or ECoG.
Editorial inference: the strong zero-shot SVM result hints that EAD embeddings separate semantic categories rather than just memorizing training classes; connecting these embeddings to image generation models would test whether the adapter preserves visual content, not only category labels.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core idea is solid: a temporal CNN adapter that distills arbitrary EEG montages into a fixed foundation-model input space, tested against a manual channel-mixing baseline. The improvements on EEG-ImageNet (99.33 vs 99.11 for their own BFM mixing) and the first subject-independent BrainLat results are plausible and worth reporting. Giving credit where due, the paper fairly compares against its own manual mixing variant and against LaBraM with raw input, which gives the reader a sense of what the adapter actually adds.

The soft spots are real but uneven. The biggest problem is the zero-shot claim in the abstract and contribution 3. Table 3 reports SVM, KNN, and K-means on six held-out classes. SVM and KNN cannot classify classes without labeled support examples from those classes; the paper never states whether such labels were used, but under the natural reading they were. So the 98.21% figure is a supervised linear probe, not zero-shot generalization. The only genuinely unsupervised result is the K-means row at 39.46%, which the text does not emphasize. That is a concrete internal inconsistency in the paper's strongest generalization claim, not a missing error bar.

The other concerns are more standard. Tables 2 through 4 report single accuracies with no standard deviations, confidence intervals, or significance tests. On a 0.22% gap, that matters. There is no code, no hyperparameter list, and the adapter architecture is underspecified ("varying number of convolution layers" without kernel sizes or channel counts). The BrainLat sample-level vs subject-level gap (80.79 vs 92.31) is large and unexplained. These are fixable but they currently block replication.

Who gets value from this paper? Researchers working on EEG foundation models or cross-dataset transfer will find the adapter concept and the channel-mixing comparison worth reading. The main classification results may well hold up. But the zero-shot claim as written is misleading, and the evaluation lacks the statistical and reproducibility details needed to take the numbers at face value.

This deserves a serious referee, not a desk reject, because the central idea is useful and the main results are potentially real. The authors need to correct the zero-shot description or remove it, add error bars, and provide a complete protocol. I would accept it for review and let the referee push for that revision.

Referee Report

3 major / 5 minor

Summary. The paper proposes EAD, a temporal-convolution adapter that maps arbitrary multichannel EEG signals into the fixed input space of the pretrained LaBraM foundation model, enabling a unified pipeline for EEG classification across different acquisition setups. The authors evaluate EAD on EEG-ImageNet and BrainLat, reporting 99.33% accuracy on EEG-ImageNet (versus 99.11% for BFM with manual channel mixing) and 92.31% subject-level accuracy on BrainLat. They also claim zero-shot generalization by holding out six EEG-ImageNet classes and reporting 98.21% accuracy with an SVM classifier. The paper positions these results as evidence of a device-agnostic, generalizable EEG feature extraction framework.

Significance. If the results hold, the adapter idea is a useful practical contribution: it would allow a single pretrained EEG foundation model to be adapted to datasets with different channel montages without hand-crafted channel alignment, and the subject-independent BrainLat result is of clinical interest. The paper reasonably builds on an external pretrained model and uses two public benchmarks with several baselines. However, the strongest advertised claim, zero-shot generalization, is not actually demonstrated as stated, and the headline accuracy differences over the strongest baseline are reported without any uncertainty quantification. The contribution is therefore incremental and currently not fully supported.

major comments (3)

[Section 4.3, Table 3, Abstract, Contribution 3] The claimed zero-shot classification is not demonstrated. The protocol described in Section 4.3 holds out six classes from training splits, but the reported SVM and KNN results require labeled examples from those six held-out classes to train or parameterize the classifiers. The manuscript never states whether labels from the six unseen classes were used to fit the SVM or KNN; under the most natural reading they were, since an SVM cannot assign one of six unseen class labels without labeled support examples. The 98.21% figure is therefore a supervised transfer or linear-probe result on frozen features, not zero-shot classification. The only unsupervised row, K-means, gives 39.46%, which is far below the EEGLSTM K-means baseline of 62.50%. The abstract and Contribution 3 claim generalization capability based on this experiment, so this is a load-bearing internal inconsistency. Either the protocol must be changed to a genuinely label-free evaluation on the six classes, or the zero-shot terminology must be removed and the supervised-transfer nature of Table 3 reported honestly.
[Tables 2, 3, and 4] All reported accuracies are single numbers with no standard deviations, confidence intervals, or significance tests. The central comparison in Table 2, EAD at 99.33% versus BFM with channel mixing at 99.11%, is a difference of 0.22 percentage points, which could easily be run-to-run noise given the lack of repeated trials. The same issue affects the BrainLat results and the SVM column of Table 3. Without multiple seeds or a statistical test, the claim that EAD outperforms the BFM mixing baseline is not established.
[Section 3.3 and Figure 2] The EAD architecture is underspecified. The text says only that a CNN with temporal convolution maps an E-channel, T-step input to the BFM's input dimensions, and that the number of convolution layers is varied per dataset. Kernel sizes, strides, channel widths, nonlinearities, adapter parameter counts, and fine-tuning hyperparameters (learning rate, epochs, batch size, optimizer, freezing strategy) are not given. This prevents reproduction and makes the claimed 'automatic distillation' difficult to evaluate, since the adapter design itself is manually adjusted per dataset. The paper should provide a complete architectural and training specification.

minor comments (5)

[Section 4.2, BrainLat paragraph] The description of windowing is incomplete: the text says a non-overlapping window of 128 timesteps is applied, but does not state the sampling rate or the resulting window duration in seconds, nor how subject-level accuracy is computed from the sample-level predictions.
[Section 4.3, first paragraph] The phrase 'in conjuction with' contains a typo; it should read 'in conjunction with'.
[Table 1 caption] The caption begins with 'T able 1', which appears to be a formatting artifact; this should be corrected to 'Table 1'.
[Section 3.2, second paragraph] The manual channel alignment procedure is described only in terms of anatomical closeness; it would clarify reproducibility to state the exact matching rule or cite a standard electrode-position mapping.
[Figure 3 and Figure 4] The t-SNE plots are mentioned as evidence of effective learning, but no interpretation is given for how the clusters support the quantitative claims; a brief reading of the cluster structure would help.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation chain; the zero-shot label is a reporting inconsistency, not a circular step.

full rationale

The derivation chain is empirical rather than deductive: EAD is a temporal CNN adapter trained end-to-end with categorical cross-entropy on two public benchmarks, on top of the externally pretrained LaBraM checkpoint (ref [18], ICLR 2024). No equation in the paper defines the reported accuracy in terms of the fitted parameters, and no fitted constant is renamed as a prediction. The 99.33% EEG-ImageNet and 92.31% BrainLat figures are direct evaluations on held-out test splits of public datasets; the BrainLat subject-level number uses a subject-independent split. The only self-citations are ref [20] (same first author, WACV 2024), used for the six-class held-out protocol and the EEGLSTM baseline, and ref [17] (coauthor A. Bhavsar), cited in related work; neither is load-bearing for the central claim. A separate reporting issue, not a circularity, is that Section 4.3 calls the six-unseen-class SVM/KNN evaluation "zero shot", but SVM and KNN require labeled examples from those classes, so the 98.21% figure is supervised transfer on frozen features; the text never states that the held-out labels are withheld from the SVM/KNN, and the unsupervised K-means row (39.46%) is the only class-zero-shot number. This weakness affects the validity of the generalization claim, but it does not make the main benchmark derivation circular, because the EAD representation was trained on the 34 training classes and the held-out labels (if used) enter only at the final SVM/KNN probe stage, not into the adapter training.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on the LaBraM feature space being universal enough for a learned channel-agnostic adapter, on hand-selected channel alignments in the manual baseline, and on several unreported tuning choices. The ledger is not heavy, but it is under-specified: the most important free parameters, the adapter architecture and training schedule, are not disclosed.

free parameters (4)

Number of EAD convolution layers = not reported
Section 3.3 states adapters are designed for each dataset by varying the number of convolution layers, so architecture depth is a per-dataset tuned choice.
Fine-tuning hyperparameters = not reported
Learning rate, epochs, batch size, optimizer, and weight decay are not given for the adapter or the BFM fine-tuning.
Manual channel mixing sets = 5 channels per BFM electrode
Table 1 maps each BFM electrode to its anatomically closest neighbors; these hand-selected groups affect the 99.11% mixing baseline and motivate the EAD design.
BrainLat window length = 128 timesteps
Section 4.2 chooses a non-overlapping window of 128 timesteps to chunk 10-minute recordings; no ablation justifies this window size.

assumptions (5)

domain assumption LaBraM's pretrained representations are meaningful generic EEG features for downstream classification.
The method fine-tunes a pretrained BFM rather than training from scratch, so the result depends on LaBraM features transferring to both stimulus and resting-state EEG.
domain assumption Anatomically close electrodes carry signal that can be mixed without losing discriminative information.
Section 3.2 and Table 1 build composite signals from adjacent electrodes; the value of this mixing is assumed from anatomical proximity.
domain assumption Channel identity and location are not needed by EAD because the temporal CNN can learn the montage mapping.
Section 3.3 feeds raw channels into a CNN with no positional or anatomical channel information, assuming the mapper can align arbitrary montages to the BFM input space.
domain assumption Excluding MS and PD subsets from BrainLat does not bias the remaining comparison.
Section 4.2 drops MS and PD due to cross-region inconsistencies, leaving only AD, bvFTD, and HC, but the paper does not analyze how this selection affects the reported accuracy.
domain assumption Non-overlapping 128-step windows of resting-state EEG are independent samples.
Section 4.2 converts each subject's 10-minute recording into many windows and treats them as separate samples, ignoring temporal autocorrelation.

invented entities (1)

EAD adapter
purpose: A temporal-convolution CNN that maps variable-channel EEG recordings into the fixed input shape of the LaBraM base foundation model.
EAD is the paper's new component, but the only evidence for it is the paper's own accuracy tables. Architecture details, layer counts, and hyperparameters are omitted, so there is no independent falsifiable handle outside the reported experiments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EAD: An EEG Adapter for Automated Classification." pith.science (2026). https://pith.science/paper/I52M2EEB

@misc{pith2026250523107,
  author       = {Pith},
  title        = {Pith review of: EAD: An EEG Adapter for Automated Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I52M2EEB}},
  note         = {Machine review of arXiv:2505.23107}
}

read the original abstract

While electroencephalography (EEG) has been a popular modality for neural decoding, it often involves task specific acquisition of the EEG data. This poses challenges for the development of a unified pipeline to learn embeddings for various EEG signal classification, which is often involved in various decoding tasks. Traditionally, EEG classification involves the step of signal preprocessing and the use of deep learning techniques, which are highly dependent on the number of EEG channels in each sample. However, the same pipeline cannot be applied even if the EEG data is collected for the same experiment but with different acquisition devices. This necessitates the development of a framework for learning EEG embeddings, which could be highly beneficial for tasks involving multiple EEG samples for the same task but with varying numbers of EEG channels. In this work, we propose EEG Adapter (EAD), a flexible framework compatible with any signal acquisition device. More specifically, we leverage a recent EEG foundational model with significant adaptations to learn robust representations from the EEG data for the classification task. We evaluate EAD on two publicly available datasets achieving state-of-the-art accuracies 99.33% and 92.31% on EEG-ImageNet and BrainLat respectively. This illustrates the effectiveness of the proposed framework across diverse EEG datasets containing two different perception tasks: stimulus and resting-state EEG signals. We also perform zero-shot EEG classification on EEG-ImageNet task to demonstrate the generalization capability of the proposed approach.

Figures

Figures reproduced from arXiv: 2505.23107 by the authors.

**Figure 1.** The figure demonstrates our approach to utilize the EAD for classification tasks. We utilize two setups resting state and image decoding multichannel EEG dataset (MCED) to showcase the effectiveness of our approach. EEG representations, which will eventually be used as a condition for various image generation techniques [17,20]. Apart from visual brain decoding tasks, EEG has numerous applications in the medical dom… view at source ↗

**Figure 2.** This image portrays a magnified view of the EAD architecture. X refers to the input EEG signal with E channels and T time steps. T ′ is dependent upon the convolution kernel, ICh is the number of input channels and IT steps is the number of time steps. LCE is categorical cross entropy loss. 4.1 Experimental Setup We first filter EEG signal with a notch filter at 50Hz and a bandpass filter with a low cutoff of 0.1Hz … view at source ↗

**Figure 3.** On the left, the tSNE plot is for the manual channel selection without mixing data. On the right, the tSNE plot is for the automatic distillation using EAD. Each color in both plots denotes a different class. This provides evidence that using EAD, effective learning is happening for the data [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: tSNE plot for BrainLat dataset [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

[1]

Brain-computer interfaces in neurological rehabilitation

Daly JJ, Wolpaw JR. Brain-computer interfaces in neurological rehabilitation. Lancet Neurol. 2008 Nov;7(11):1032-43. doi: 10.1016/S1474-4422(08)70223-0. Epub 2008 Oct 2. PMID: 18835541

doi:10.1016/s1474-4422(08)70223-0 2008
[2]

Van der Maaten, L., & Hinton, G. (2008). Visualizing data using t-SNE.Journal of machine learning research,9(11)

work page 2008
[3]

Noninvasive Electroencephalo- gram Based Control of a Robotic Arm for Reach and Grasp Tasks

Meng J, Zhang S, Bekyo A, Olsoe J, Baxter B, He B. Noninvasive Electroencephalo- gram Based Control of a Robotic Arm for Reach and Grasp Tasks. Sci Rep. 2016 Dec 14;6:38565. doi: 10.1038/srep38565. Erratum in: Sci Rep. 2020 Apr 15;10(1):6627. doi: 10.1038/s41598-020-63070-z. PMID: 27966546; PMCID: PMC5155290

work page doi:10.1038/srep38565 2016
[4]

Obeid, I., & Picone, J. (2016). The temple university hospital EEG data corpus. Frontiers in neuroscience,10, 196

work page 2016
[5]

Spampinato, C., Palazzo, S., Kavasidis, I., Giordano, D., Souly, N., & Shah, M. (2017). Deep learning human mind for automated visual classification. InProceed- ings of the IEEE conference on computer vision and pattern recognition(pp. 6809- 6817). 6.https://github.com/perceivelab/eeg_visual_classification/issues/4

work page 2017
[7]

(2017, Octo- ber)

Kavasidis, I., Palazzo, S., Spampinato, C., Giordano, D., & Shah, M. (2017, Octo- ber). Brain2image: Converting brain signals into images. InProceedings of the 25th ACM international conference on Multimedia(pp. 1809-1817)

work page 2017
[8]

Palazzo, S., Spampinato, C., Kavasidis, I., Giordano, D., & Shah, M. (2017). Gener- ative adversarial networks conditioned by brain signals. InProceedings of the IEEE international conference on computer vision(pp. 3410-3418)

work page 2017
[9]

Van Den Oord, A., & Vinyals, O. (2017). Neural discrete representation learning. Advances in neural information processing systems,30

work page 2017

Show all 30 references

[10]

Cassani, Raymundo, Estarellas, Mar, San-Martin, Rodrigo, Fraga, Francisco J., Falk, Tiago H., Systematic Review on Resting-State EEG for Alzheimer’s Disease Diagnosis and Progression Assessment,Disease Markers, 2018, 5174815, 26 pages,

2018
[11]

P., Smith, L

Strein, M., Holton-Burke, J. P., Smith, L. R., & Brophy, G. M. (2019). Prevention, Treatment, and Monitoring of Seizures in the Intensive Care Unit.Journal of clinical medicine,8(8), 1177. https://doi.org/10.3390/jcm8081177

2019 doi
[12]

Jiang, J., Fares, A., & Zhong, S. H. (2019). A context-supported deep learn- ing framework for multimodal brain imaging classification.IEEE Transactions on Human-Machine Systems,49(6), 611-622

2019
[13]

Jiang, J., Fares, A., & Zhong, S. H. (2020). A brain-media deep framework towards seeing imaginations inside brains.IEEE Transactions on Multimedia,23, 1454-1465

2020
[14]

Palazzo, S., Spampinato, C., Kavasidis, I., Giordano, D., Schmidt, J., & Shah, M. (2020). Decoding brain representations by multimodal learning of neural activity and visual features.IEEE Transactions on Pattern Analysis and Machine Intelli- gence,43(11), 3833-3849

2020
[15]

N., Amar, L., & Udutalapalli, V

Khare, S., Choubey, R. N., Amar, L., & Udutalapalli, V. (2022). Neurovision: perceived image regeneration using cprogan.Neural Computing and Applications, 34(8), 5979-5991

2022
[16]

& Ibañez, A

Prado, P., Medel, V., Gonzalez-Gomez, R., Sainz-Ballesteros, A., Vidal, V., Santamaría-García, H., ... & Ibañez, A. (2023). The BrainLat project, a multimodal neuroimaging dataset of neurodegeneration from underrepresented backgrounds. Scientific Data,10(1), 889

2023
[17]

R., & Bhavsar, A

Mishra, R., Sharma, K., Jha, R. R., & Bhavsar, A. (2023). NeuroGAN: image reconstruction from EEG signals via an attention-based GAN.Neural Computing and Applications,35(12), 9181-9192

2023
[18]

Jiang, W., Zhao, L., & Lu, B. L. Large Brain Model for Learning Generic Rep- resentations with Tremendous EEG Data in BCI. InThe Twelfth International Conference on Learning Representations
[19]

(2024, March)

Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., & Shan, Y. (2024, March). T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. InProceedings of the AAAI conference on artificial intelligence (Vol. 38, No. 5, pp. 4296-4304)

2024
[20]

Singh, P., Dalal, D., Vashishtha, G., Miyapuram, K., & Raman, S. (2024). Learn- ing robust deep visual representations from eeg brain recordings. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision(pp. 7553- 7562)

2024
[21]

A novel hybrid model in the diagnosis and classification of alzheimer’s disease using eeg signals: Deep ensemble learning (del) approach

Nour, M., Senturk, U., and Polat, K. A novel hybrid model in the diagnosis and classification of alzheimer’s disease using eeg signals: Deep ensemble learning (del) approach. Biomedical Signal Processing and Control, 89:105751, 2024

2024
[22]

and Saravanan, S

kumar Ravikanti, D. and Saravanan, S. Eegalzheimer’snet: Development of transformer-based attention long short term memory network for detecting alzheimer disease using eeg signal. Biomedical Signal Processing and Control, 86:105318, 2023

2023
[23]

A deep learning model for the detection of vari ous dementia and mci pathologies based on resting-state electroencephalog- raphy data: A retrospective multicentre study

Watanabe, Y., Miyazaki, Y., Hata, M., Fukuma, R., Aoki, Y., Kazui, H., Araki, T., Taomoto, D., Satake, Y., Suehiro, T., et al. A deep learning model for the detection of vari ous dementia and mci pathologies based on resting-state electroencephalog- raphy data: A retrospective...

2024
[24]

A multi- modal classification method for early diagnosis of mild cognitive impairment and alzheimer’s disease using three paradigms with various task difficulties

Chen, S., Zhang, C., Yang, H., Peng, L., Xie, H., Lv, Z., and Hou, Z.-G. A multi- modal classification method for early diagnosis of mild cognitive impairment and alzheimer’s disease using three paradigms with various task difficulties. IEEE Trans- actions on Neural Systems an...

2024
[25]

Boonyakitanont, P., Lek-Uthai, A., Chomtho, K., Songsiri, J. (2020). A review of feature extraction and performance evaluation in epileptic seizure detection using EEG. Biomedical Signal Processing and Control, 57, 101702. EAD: An EEG Adapter for Automated Classification 11

2020
[26]

U., Alsulaiman, M., Muhammad, G., Mekhtiche, M

Amin, S. U., Alsulaiman, M., Muhammad, G., Mekhtiche, M. A., Hossain, M. S. (2019). Deep Learning for EEG motor imagery classification based on multi-layer CNNs feature fusion. Future Generation computer systems, 101, 542-554

2019
[27]

S., Mountstephens, J., Teo, J

Suhaimi, N. S., Mountstephens, J., Teo, J. (2020). EEG-based emotion recogni- tion: a state-of-the-art review of current trends and opportunities. Computational intelligence and neuroscience, 2020(1), 8875426

2020
[28]

IEEE transactions on neural systems and rehabilitation engineering, 25(5), 402-412

Biesmans,W.,Das,N.,Francart,T.,Bertrand,A.(2016).Auditory-inspiredspeech envelope extraction methods for improved EEG-based auditory attention detection in a cocktail party scenario. IEEE transactions on neural systems and rehabilitation engineering, 25(5), 402-412

2016
[29]

K., & Roy, P

Mukherjee, P., Das, A., Bhunia, A. K., & Roy, P. P. (2019, September). Cogni-net: Cognitive feature learning through deep visual perception. In2019 IEEE Interna- tional Conference on Image Processing (ICIP)(pp. 4539-4543). IEEE

2019
[30]

A., Dumais, S

Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J., & Scholkopf, B. (1998). Support vector machines. IEEE Intelligent Systems and their applications, 13(4), 18-28

1998
[2018]

Singh et al

https://doi.org/10.1155/2018/5174815 10 P. Singh et al

2018 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Brain-computer interfaces in neurological rehabilitation

Daly JJ, Wolpaw JR. Brain-computer interfaces in neurological rehabilitation. Lancet Neurol. 2008 Nov;7(11):1032-43. doi: 10.1016/S1474-4422(08)70223-0. Epub 2008 Oct 2. PMID: 18835541

doi:10.1016/s1474-4422(08)70223-0 2008

[2] [2]

Van der Maaten, L., & Hinton, G. (2008). Visualizing data using t-SNE.Journal of machine learning research,9(11)

work page 2008

[3] [3]

Noninvasive Electroencephalo- gram Based Control of a Robotic Arm for Reach and Grasp Tasks

Meng J, Zhang S, Bekyo A, Olsoe J, Baxter B, He B. Noninvasive Electroencephalo- gram Based Control of a Robotic Arm for Reach and Grasp Tasks. Sci Rep. 2016 Dec 14;6:38565. doi: 10.1038/srep38565. Erratum in: Sci Rep. 2020 Apr 15;10(1):6627. doi: 10.1038/s41598-020-63070-z. PMID: 27966546; PMCID: PMC5155290

work page doi:10.1038/srep38565 2016

[4] [4]

Obeid, I., & Picone, J. (2016). The temple university hospital EEG data corpus. Frontiers in neuroscience,10, 196

work page 2016

[5] [5]

Spampinato, C., Palazzo, S., Kavasidis, I., Giordano, D., Souly, N., & Shah, M. (2017). Deep learning human mind for automated visual classification. InProceed- ings of the IEEE conference on computer vision and pattern recognition(pp. 6809- 6817). 6.https://github.com/perceivelab/eeg_visual_classification/issues/4

work page 2017

[6] [7]

(2017, Octo- ber)

Kavasidis, I., Palazzo, S., Spampinato, C., Giordano, D., & Shah, M. (2017, Octo- ber). Brain2image: Converting brain signals into images. InProceedings of the 25th ACM international conference on Multimedia(pp. 1809-1817)

work page 2017

[7] [8]

Palazzo, S., Spampinato, C., Kavasidis, I., Giordano, D., & Shah, M. (2017). Gener- ative adversarial networks conditioned by brain signals. InProceedings of the IEEE international conference on computer vision(pp. 3410-3418)

work page 2017

[8] [9]

Van Den Oord, A., & Vinyals, O. (2017). Neural discrete representation learning. Advances in neural information processing systems,30

work page 2017