The DOI matches its Crossref record; 114 of the 137 references it cites were found in a registry or an index; 40 of the 58 sources its claims cite were checked for support.
IENet: a robust convolutional neural network for EEG based brain-computer interfaces
Open accessAbstract
Abstract Objective.Brain-computer interfaces (BCIs) based on electroencephalogram (EEG) develop into novel application areas with more complex scenarios, which put forward higher requirements for the robustness of EEG signal processing algorithms. Deep learning can automatically extract discriminative features and potential dependencies via deep structures, demonstrating strong analytical capabilities in numerous domains such as computer vision and natural language processing. Making full use of deep learning technology to design a robust algorithm that is capable of analyzing EEG across BCI paradigms is our main work in this paper.Approach.Inspired by InceptionV4 and InceptionTime architecture, we introduce a neural network ensemble named InceptionEEG-Net (IENet), where multi-scale convolutional layer and convolution of length 1 enable model to extract rich high-dimensional features with limited parameters. In addition, we propose the average receptive field (RF) gain for convolutional neural networks (CNNs), which optimizes IENet to detect long patterns at a smaller cost. We compare with the current state-of-the-art methods across five EEG-BCI paradigms: steady-state visual evoked potentials (VEPs), epilepsy EEG, overt attention P300 VEPs, covert attention P300 visual-EPs and movement-related cortical potentials.Main results.The classification results show that the generalizability of IENet is on par with the state-of-the-art paradigm-agnostic models on test datasets. Furthermore, the feature explainability analysis of IENet illustrates its capability to extract neurophysiologically interpretable features for different BCI paradigms, ensuring the reliability of algorithm.Significance.It can be seen from our results that IENet can generalize to different BCI paradigms. And it is essential for deep CNNs to increase the RF size using average RF gain.
Overview
CoChat Summary
TL;DR
IENet is a CNN ensemble combining multi-scale convolutions and length-1 convolutions, optimized with an average receptive field gain metric, for EEG-based BCI across paradigms. It performs comparably to state-of-the-art paradigm-agnostic models on five EEG datasets and extracts neurophysiologically interpretable features.
Key findings
- On SSVEP and epilepsy EEG, IENet significantly outperformed DeepConvNet and EEGNet (p<0.05) for both single models and ensembles; on overt and covert attention P300, EEGNet was best for a single model (p<0.05) but ensemble differences were not significant (p>0.05); on MRCP, IENet ranked third for a single model and second for the ensemble (p<0.05).
- Ensemble gains for IENet were 7.93% on SSVEP, 2.42% on epilepsy EEG, 2.53% on overt attention P300, 3.04% on covert attention P300, and 4.68% on MRCP.
- The optimized IE model (RF size 229) outperformed the original IE model (RF 169) and deepened IE model (RF 225) across all datasets: SSVEP 86.52% vs 82.36% vs 80.97%; epilepsy EEG 81.74% vs 80.21% vs 76.52%; overt attention P300 91.74% vs 89.29% vs 87.06%; covert attention P300 75.90% vs 70.82% vs 74.77%; MRCP 87.28% vs 83.97% vs 81.82%.
- Feature explainability showed long convolutions contributed the largest proportion across depths; SSVEP temporal kernels matched the stimulus frequency range 9.25–14.75 Hz; P300 peak positive potential appeared at approximately 330 ms for overt attention and 550 ms for covert attention; MRCP relevance was highest approximately 50 ms after right-hand movement.
Methods
- IENet is an ensemble of three IE models; each IE model contains six IE modules with multi-scale convolutional layers using long, medium, and short convolutions, plus FdrConv1, HdeConv1, CieConv1, residual connections, batch normalization, and global average pooling. The average RF gain metric guided increasing the long convolution length to 20.
- Five datasets were used: SSVEP (10 subjects, 8 channels, 12 classes, 7200 epochs), epilepsy EEG (10 subjects, 1 channel, 5 classes, 11,500 epochs), overt attention P300 (10 subjects, 16 channels, 2 classes, 11,590 epochs), covert attention P300 (same), and MRCP (12 subjects, 32 channels, 6 labels, 11,883,171 epochs). Cross-validation included leave-one-subject-out for SSVEP and P300, 25-fold for epilepsy EEG, and 8-fold for MRCP; metrics were ACC and AUC, with one-way ANOVA, Friedman, and Wilcoxon tests.
- Comparisons were made against EEGNet and DeepConvNet. Feature explainability used trained kernel weights, hidden-unit activations, t-SNE, Morlet wavelet transform, and four DeepExplain gradient-based methods; ablation studies examined convolutions, residual connections, batch normalization, and ensemble size.
Limitations
- The paper states that IENet’s test-set accuracy during training was not stable, that it was sensitive to SSVEP stimulus onset, and that a large discrepancy remained between IENet and SOTA results on SSVEP and MRCP, with a mean accuracy/AUC difference of about 10%.
- The paper states that future work is needed to test more EEG-BCI paradigms and deploy IENet in a real BCI system; the data availability statement says no new data were created or analyzed.
- Evident from the text: the three-model ensemble increases model size and computation, though the authors chose three to balance accuracy and response speed; evaluation is offline on existing datasets rather than online or real-time BCI use.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
34 Valid · 3 Not valid · 10 Not checked · 130 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“Therefore, we apply the ensemble of three IE models instead of a single one to improve stability [62].”
What the sources say
- Supportedcites [62] Fawaz, 2020, Inceptiontime: finding alexnet for time series classification, Data Min. Knowl. Discovery source
Section 3.2 (p. 7) says a single Inception network had high variability and that the authors used ensembling to address it. The citing authors' choice of three models is their own; the source used five, but supports the cited rationale for ensembling instead of using a single model. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Additionally, data-driven neural networks are required to generalize to the variable size of EEG datasets due to the different acquisition conditions of BCI paradigms [13].”
What the sources say
- Supportedcites [13] Lawhern, 2018, EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces, J. Neural Eng. source
The Introduction (p. 2) says training-data amounts varied substantially across studies, partly because of difficulties collecting data under different experimental designs, and identifies generalization to variable training-data sizes as an open question. The Discussion (p. 19) reports that EEGNet generalized across BCI paradigms with different dataset sizes. (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The difficulty of EEG signal processing is to extract high-dimensional features from non-stationary time series with low signal-to-noise ratio [18].”
What the sources say
- Partly supportedcites [18] Craik, 2019, Deep learning for electroencephalogram (EEG) classification tasks: a review, J. Neural Eng. sourceExperimental — accuracy not yet measured
The Introduction describes EEG data as high-dimensional, typically having low signal-to-noise ratios, and identifies feature extraction as part of the classification pipeline. Section 3.3 also discusses the challenge of formulating inputs and extracting EEG features. Our copy does not establish the claim’s specific description of EEG as non-stationary time series. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Specifically, with the increasing implementation of EEG-BCI, the algorithm design relied on expert experience is time-consuming and difficult [6].”
What the sources say
- Supportedcites [6] Zhang, 2021, A survey on deep learning-based non-invasive brain signals: recent advances and new frontiers, J. Neural Eng. source
Section 1.1 (General workflow; page unavailable) says EEG feature extraction depends heavily on domain knowledge and that manual extraction is time-consuming and difficult. Section 1.2 (Why deep learning?; page unavailable) likewise describes its dependence on human expertise. These passages support the citing sentence’s point about expert-dependent algorithm design for EEG-BCI. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“where multi-scale convolutional layer and convolution of length 1 enable model to extract rich high-dimensional features with limited parameters.”
- Also stated in: § IENet: a robust convolutional neural network for EEG based brain-computer interfaces · p. 4 · plain
“multi-scale convolutional layers (MSC Layers) [19, 62] and convolutions of length 1 (Conv1) [63] are adopted to enable IENet to extract rich high-dimensional features with limited parameters.”
What the sources say
- Partly supportedcites [19] Szegedy, 2017, Inception-v4, inception-resnet and the impact of residual connections on learning sourceExperimental — accuracy not yet measured
The Residual Inception Blocks section describes a 1 × 1 convolution that expands filter-bank dimensionality, and the Related Work section says a 1 × 1 convolution can reduce computational cost (page numbers unavailable). The paper also says its Inception models keep parameter counts in check, but it does not evaluate IENet or establish that these layers enable IENet to extract rich features with limited parameters. (LLM, conf=0.82)
- Partly supportedcites [62] Fawaz, 2020, Inceptiontime: finding alexnet for time series classification, Data Min. Knowl. Discovery sourceExperimental — accuracy not yet measured
Section 3.1 (p. 6) says filters of different lengths extract multiresolution features, while the length-1 bottleneck reduces dimensionality and permits longer filters with roughly the same number of parameters. The paper establishes this for InceptionTime, not specifically for IENet; it therefore supports the architectural rationale but not its application to IENet. (LLM, conf=0.84)
- Partly supportedcites [63] Lin, 2013, Network in network sourceExperimental — accuracy not yet measured
Section 3.1 (p. 3) describes 1×1 convolution as part of NIN and says its cross-channel operations allow complex, learnable feature interactions. The paper does not describe a multi-scale convolutional layer or say that these layers extract features with limited parameters; its parameter-free claim concerns global average pooling (Section 3.2, p. 4). It studies NIN, not IENet. (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § IENet: a robust convolutional neural network for EEG based brain-computer interfaces · p. 4 · plain
96 more claims · 7 with their cited sources checked
Create a free account to see all 177 claims and to verify the 3 references still open.
- It may state two separate claims as one. 48
- Probably not a substantial claim. 39
- The quote only partly states this claim. 1
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2.1.2. Average RF gain · p. 8
“Specifically, the lengths of the long, medium and short convolutions in the MSC Layer of our original designed model are 5, 10, and 15 respectively.”
§ 5.3. Effect of average RF gain · p. 23
“The length of the long, medium and short convolutions of the MSC Layer in the original IE model are 15, 10, and 5 respectively.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2.2. State-of-the-art algorithms used for comparison · p. 9
“| SSVEP | SSEP | 250 | 8 | 12 classes | 7200 | 10 |”
§ 3.1.1. Dateset 1: SSVEP · p. 9
“All data was down-sampled to 256 Hz, and then filtered from 6 to 80 Hz using an infinite impulse response filter after acquisition [64].”
Cited works that were retracted
Works this paper cites that were later retracted, from the retraction notices publishers deposit. Major when the notice came out before this paper was published.
A note, not an issue: for example, a cited work that was retracted only after this paper was published, which its authors could not have known.
In the paper
“Liu, 2021, Multi-class motor imagery EEG classification method with high accuracy and low individual differences based on hybrid neural network, J. Neural Eng., 18, 0460f1, 10.1088/1741-2552/ac1ed0”
Cited work 10.1088/1741-2552/ac1ed0 · retracted 2023-03-22 · notice 10.1088/1741-2552/acc172 · this paper published 2022-06-01
Citations
Cites
1 retracted
Liu, J., Ye, F., & Xiong, H. (2021) · Journal of Neural Engineering · doi:10.1088/1741-2552/ac1ed0 · Cited by 25
Self-citationRetractedDeep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 3×
“From AlexNet [70] to VGG [46], to ResNet [20] and Inception [19], the development of CNNs in CV also corresponds to the improvement of the RF [71].”
MethodResultVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 5×
“The SOTA model for epilepsy EEG dataset is fine-tuning VGG16 (FT-VGG16), which converts the raw EEG signals to time-frequency images as input through continuous wavelet transform.”
MethodResultInfluentialImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 1×
“From AlexNet [70] to VGG [46], to ResNet [20] and Inception [19], the development of CNNs in CV also corresponds to the improvement of the RF [71].”
ResultAdam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946 · cited 1×
“Each model is trained using the Adam optimization algorithm [90] and model weights are initialized by Glorot’s uniform technique [91].”
Method(2008) · Journal of Machine Learning Research · Cited by 33,051
Rethinking the Inception Architecture for Computer Vision
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138 · cited 1×
“Like Inception [60] in the CV field and transformer [61] in the natural language processing (NLP) field, these deep learning models have strong capability of extracting discriminative features for different tasks.”
BackgroundBatch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Ioffe, S., & Szegedy, C. (2015) · arXiv · doi:10.48550/arxiv.1502.03167 · Cited by 27,818 · cited 1×
“We adopt batch normalization [67] as the regularization method in IE module.”
Method(2010) · Cited by 14,134
EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces
Lawhern, V. J., Solon, A. J., Waytowich, N. R., Gordon, S. M., Hung, C. P., & Lance, B. J. (2018) · Journal of Neural Engineering · doi:10.1088/1741-2552/aace8c · Cited by 5,115 · cited 23×
“And they perform better than EEGNet ensemble according to the critical difference diagram.”
BackgroundMethodResultInfluential
Cited by
A transformer-based deep neural network model for SSVEP classification
Chen, J., Zhang, Y., Pan, Y., Xu, P., & Guan, C. (2023) · Neural Networks · doi:10.1016/j.neunet.2023.04.045 · Cited by 170
Status of deep learning for EEG-based brain–computer interface applications
Hossain, K. M., Islam, M. A., Hossain, S., Nijholt, A., & Ahad, M. A. R. (2023) · Frontiers in Computational Neuroscience · doi:10.3389/fncom.2022.1006763 · Cited by 188
EEG-based emotion recognition using multi-scale dynamic CNN and gated transformer
Cheng, Z., Bu, X., Wang, Q., Yang, T., & Tu, J. (2024) · Scientific Reports · doi:10.1038/s41598-024-82705-z · Cited by 128
Elashmawi, W. H., Ayman, A., Antoun, M., Mohamed, H. C., Mohamed, S. E., Amr, H., Talaat, Y., & Ali, A. (2024) · Applied Sciences · doi:10.3390/app14146347 · Cited by 97
Ahn, H.-J., Lee, D.-H., Jeong, J.-H., & Lee, S. (2022) · IEEE Transactions on Neural Systems and Rehabilitation Engineering · doi:10.1109/tnsre.2022.3229330 · Cited by 82
“Various kinds of deep learning models were proposed to classify EEG signals [11], [12], [13], [14].”
Background
Metadata
- Authors
- Du, Y., & Liu, J.
- Affiliations
- University of Science and Technology Beijing (CN)
- Corresponding author
- Liu, Jian
- ORCID
- Venue
- Journal of Neural Engineering
- Publisher
- IOP Publishing
- Published
- 2022-06-01
- Volume
- 19
- Issue
- 3
- Pages
- 036031
- ISSN
- 1741-2552 (ISSN-L)
- 1741-2560
- OpenAlex
- W4281399130
- Cited by
- 34
- References
- 137
- Citations by year
- 2026: 5 · 2025: 9 · 2024: 14 · 2023: 5 · 2022: 1
- FWCI
- 3.74
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (vor, 6 d)
- iopscience.iop.org/info/page/text-and-data-mining (tdm, 6 d)
- Locations
- Journal of Neural Engineering (publishedVersion, cc-by)
- PubMed (publishedVersion)
- Topics
- Life Sciences › Neuroscience › Cognitive Neuroscience › EEG and Brain-Computer Interfaces
- Physical Sciences › Computer Science › Artificial Intelligence › Emotion and Mood Recognition
- Physical Sciences › Computer Science › Human-Computer Interaction › Gaze Tracking and Assistive Technology
- Keywords
- EEG-based brain-computer interface, neural network ensembles, Inception-v4, InceptionTime, multi-scale convolution, steady-state visual evoked potentials, P300, covert attention, movement-related cortical potentials
- MeSH terms
- Algorithms, Electroencephalography, Evoked Potentials, Visual, Reproducibility of Results, Neural Networks, Computer, Brain-Computer Interfaces
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: IENet: a robust convolutional neural network for EEG based brain-computer interfaces. https://app.cochat.ai/papers/10.1088/1741-2552/ac7257
- The paper
- Du, Y., & Liu, J. (2022). IENet: a robust convolutional neural network for EEG based brain-computer interfaces. Journal of Neural Engineering. https://doi.org/10.1088/1741-2552/ac7257
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.