CoChat

Motor imagery EEG classification via wavelet-packet synthetic augmentation and entropy-based channel selection

Open access
Zheng, Minmin, Qian, Zhengkang, Zhao, Tong · 2025 · Frontiers in Neuroscience

The DOI matches its Crossref record; 45 of the 46 references it cites were found in a registry or an index, and 1 are not checked yet; 1 of the 2 sources its claims cite were checked for support.

Abstract

Introduction: Motor-imagery (MI) brain-computer interfaces often suffer from limited EEG datasets and redundant channels, hampering both accuracy and clinical usability. We address these bottlenecks by presenting a unified framework that simultaneously boosts classification performance, reduces the number of required sensors, and eliminates the need for extra recordings. Methods: A three-stage pipeline is proposed. (1) Wavelet-packet decomposition (WPD) partitions each MI class into low-variance "stable" and high-variance "variant" trials; sub-band swapping between matched pairs generates synthetic trials that preserve event-related desynchronization/synchronization signatures. (2) Channel selection uses wavelet-packet energy entropy (WPEE) to quantify both spectral-energy complexity and class-separability; the top-ranked leads are retained. (3) A lightweight multi-branch network extracts multi-scale temporal features through parallel dilated convolutions, refines spatial patterns via depth-wise convolutions, and feeds the fused spatiotemporal tensor to a Transformer encoder with multi-head self-attention; soft-voted fully-connected layers deliver robust class labels. Results: < 0.01). Discussion: Integrating WPD-based augmentation with WPEE-driven channel selection yields higher MI decoding accuracy with fewer channels and without extra recordings. The framework offers a computationally efficient, clinically viable paradigm for enhanced EEG classification in resource-constrained settings.

Verified 2026-10-10 · 1 reference not checked yet

Overview

CoChat Summary

TL;DR

The paper proposes DACS-TBTrans, a three-stage motor-imagery EEG framework combining wavelet-packet synthetic augmentation, wavelet-packet energy-entropy channel selection, and a multi-branch CNN–Transformer with voting. It reports 86.81% and 86.64% mean accuracy on BCI Competition IV 2a and PhysioNet MI while removing 27% of sensors.

Key findings

  • On BCI Competition IV 2a, DACS-TBTrans achieved 86.81% mean accuracy and κ = 0.8125, outperforming baselines including TCNet fusion (83.73%), M-FANet (79.28%), and conformer (78.66%).
  • On PhysioNet MI, it achieved 86.64% mean accuracy and F1 = 0.86, with peak subject accuracy of 97.61%, using 103 valid subjects after excluding 6.
  • The method removed 27% of sensors and outperformed the same network trained on all 22 channels, with paired t-tests reported as significant (p < 0.01).
  • Ablation on BCI IV 2a: channel selection alone improved average accuracy by 1.74%, data augmentation alone by 7.33%, and the combined DACS approach by 8.92% over the baseline TBTrans.

Methods

  • Three-stage pipeline: (1) wavelet-packet decomposition with α/β sub-band swapping between low- and high-variance same-class trials for synthetic training data; (2) channel selection using wavelet-packet energy entropy differences, with a 12 + 4 symmetric configuration and n = 16 channels; (3) a three-branch temporal convolutional network with spatial convolutions, Transformer encoders, and soft-voted fully connected layers.
  • Data: BCI Competition IV 2a (9 subjects, 4 MI classes, 288 trials/subject) and PhysioNet MI (109 subjects, 64 electrodes at 160 Hz; 103 subjects retained after exclusions). Preprocessing used 1 Hz high-pass, 50 Hz notch, 4–40 Hz band-pass, and overlapping 4-s windows from 2–6 s and 3–7 s post-cue.
  • Evaluation: accuracy, Kappa, and F1; paired t-tests, Wilcoxon signed-rank tests with Holm–Bonferroni correction, Friedman ranking, ablation, window-length tests, and parameter/latency comparison.

Limitations

  • The paper states the filter-only preprocessing cannot remove ocular, cardiac, or muscular artifacts whose spectra overlap α/β bands, so within-subject results may be slightly inflated; future work will explore ICA or ASR.
  • The learning curve for Subject 3 shows training accuracy near 100% while testing accuracy stabilizes around 88%, with testing loss fluctuating and remaining relatively high, indicating a generalization gap and potential overfitting.
  • The paper notes misclassification between tongue and feet in some subjects may be due to suboptimal channel selection, and future work should adapt channel selection to inter-subject variability. Evident from the text: the channel-selection description first says n = 12 is later verified, but the parameter experiment sets n = 16, an inconsistency in the reported hyperparameter.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fnins.2025.1689647

Cites

45 of 46 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Flowchart of the DACS-TBTrans framework.
Figure 1 Flowchart of the DACS-TBTrans framework.CC BY · PubMed Central
Figure 2 Flowchart of the data augmentation.
Figure 2 Flowchart of the data augmentation.CC BY · PubMed Central
Figure 3 Comparative classification accuracy of proposed method vs. baselines on BCI competition IV 2a dataset.
Figure 3 Comparative classification accuracy of proposed method vs. baselines on BCI competition IV 2a dataset.CC BY · PubMed Central
Figure 4 Model performance comparison with statistical significance.
Figure 4 Model performance comparison with statistical significance.CC BY · PubMed Central
Figure 5 Model ranking based on Friedman test.
Figure 5 Model ranking based on Friedman test.CC BY · PubMed Central
Figure 6 Confusion matrix of proposed method on BCI competition IV 2a dataset.
Figure 6 Confusion matrix of proposed method on BCI competition IV 2a dataset.CC BY · PubMed Central
Figure 7 Classification accuracy of each subject in the Physionet MI dataset.
Figure 7 Classification accuracy of each subject in the Physionet MI dataset.CC BY · PubMed Central
Figure 8 Confusion matrix of proposed method on Physionet MI dataset.
Figure 8 Confusion matrix of proposed method on Physionet MI dataset.CC BY · PubMed Central
Figure 9 Feature visualization of proposed method for Subject 1 on BCI competition IV 2a dataset.
Figure 9 Feature visualization of proposed method for Subject 1 on BCI competition IV 2a dataset.CC BY · PubMed Central
Figure 10 Ablation results of proposed method on BCI competition IV 2a dataset.
Figure 10 Ablation results of proposed method on BCI competition IV 2a dataset.CC BY · PubMed Central
Figure 11 Learning curve of subject 3 on BCI competition IV 2a dataset.
Figure 11 Learning curve of subject 3 on BCI competition IV 2a dataset.CC BY · PubMed Central
Figure 12 Visualization of brain topography maps of some subjects on the BCI Competition IV 2a dataset.
Figure 12 Visualization of brain topography maps of some subjects on the BCI Competition IV 2a dataset.CC BY · PubMed Central
Figure 13 Channel selection hyperparameter optimization results.
Figure 13 Channel selection hyperparameter optimization results.CC BY · PubMed Central
Figure 14 Channel symmetry comparison: Subjects 1, 3, 7 vs. Subjects 2, 5, 6.
Figure 14 Channel symmetry comparison: Subjects 1, 3, 7 vs. Subjects 2, 5, 6.CC BY · PubMed Central
Figure 15 Channel selection results with strategic supplementation.
Figure 15 Channel selection results with strategic supplementation.CC BY · PubMed Central

Verification

Claims the paper makes

1 Valid · 1 Not valid · 1 Not checked · 81 Own results and interpretations

34 more claims

Create a free account to see all 84 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 38
  • Probably not a substantial claim. 9

Create a free account to see these claims.

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.

5 of 9 · the rest aren’t loaded

4 more findings

Create a free account to see every finding of this check.

See all findings — free account

Citations

Cites

Show all 46 references

Research works printed but not linked: 2

  1. Song, 2021, arXiv [preprint], 2102.04456
  2. Xiong, 2021, Control of upper limb rehabilitation robot based on steady-state visual evoked potential, Sci. Technol. Eng., 21, 7237, 10.3969/j.issn.1671-1815.2021.17.037 · doi:10.3969/j.issn.1671-1815.2021.17.037

Metadata

Authors
Zheng, M., Qian, Z., & Zhao, T.
Affiliations
Putian University (CN)
Corresponding author
Zheng, Minmin
Venue
Frontiers in Neuroscience
Published
2025-11-10
Volume
19
Pages
1689647-1689647
ISSN
1662-453X (ISSN-L)
OpenAlex
W4416096269
Cited by
1
References
35
Citations by year
2026: 1
FWCI
0.57
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Neuroscience (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Fujian Provincial Department of Science and Technology
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › EEG and Brain-Computer Interfaces
  • Physical Sciences › Engineering › Biomedical Engineering › Muscle activation and electromyography studies
  • Physical Sciences › Computer Science › Artificial Intelligence › Emotion and Mood Recognition
Keywords
motor imagery, EEG classification, brain-computer interface, wavelet packet decomposition, synthetic data augmentation, wavelet packet energy entropy, channel selection, Transformer encoder, event-related desynchronization, event-related synchronization, soft voting, resource-limited settings

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.

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: Motor imagery EEG classification via wavelet-packet synthetic augmentation and entropy-based channel selection. https://app.cochat.ai/papers/10.3389/fnins.2025.1689647
The paper
Zheng, M., Qian, Z., & Zhao, T. (2025). Motor imagery EEG classification via wavelet-packet synthetic augmentation and entropy-based channel selection. Frontiers in Neuroscience. https://doi.org/10.3389/fnins.2025.1689647

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.