Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 46 of the 48 references it cites were found in a registry or an index; 4 of the 13 sources its claims cite were checked for support.

EEG-based emotion recognition with autoencoder feature fusion and MSC-TimesNet model

Yin, Jibin, Qiao, Zhijian, Han, Luyao, Zhang, Xiangliang · 2026 · Computer Methods in Biomechanics and Biomedical Engineering

Abstract

Electroencephalography (EEG) signals are widely employed due to their spontaneity and robustness against artifacts in emotion recognition. However, existing methods are often unable to fully integrate high-dimensional features and capture changing patterns in time series when processing EEG signals, which results in limited classification performance. This paper proposes an emotion recognition method (AEF-DL) based on autoencoder fusion features and MSC-TimesNet models. Firstly, we segment the EEG signal in five frequency bands into time windows of 0.5 s, extract power spectral density (PSD) features and differential entropy (DE) features, and implement feature fusion using the autoencoder to enhance feature representation. Based on the TimesNet model and incorporating the multi-scale convolutional kernels, this paper proposes an innovative deep learning model (MSC-TimesNet) for processing fused features. MSC-TimesNet efficiently extracts inter-period and intra-period information. To validate the performance of the proposed method, we conducted systematic experiments on the public datasets DEAP and Dreamer. In dependent experiments with subjects, the classification accuracies reached 98.97% and 95.71%, respectively; in independent experiments with subjects, the accuracies reached 97.23% and 92.95%, respectively. These results demonstrate that the proposed method exhibits significant advantages over existing methods, highlighting its effectiveness and broad applicability in emotion recognition tasks.

Overview

CoChat Summary

TL;DR

The paper proposes AEF-DL, an EEG emotion recognition pipeline that fuses power spectral density (PSD) and differential entropy (DE) features with a denoising autoencoder and classifies them with MSC-TimesNet, a TimesNet variant augmented with multi-scale convolutional kernels. It reports very high arousal/valence classification accuracies on DEAP and DREAMER in both subject-dependent and subject-independent settings.

Key findings

  • Subject-dependent experiments: average accuracies on DEAP were 98.97% (arousal) and 98.90% (valence), with average F scores of 98.99% and 98.97%; on DREAMER, 95.71% (arousal) and 95.86% (valence), with F scores of 94.48% and 94.37%.
  • Subject-independent experiments: MSC-TimesNet reached 97.23% (arousal) and 97.02% (valence) on DEAP, and 92.95% (arousal) and 93.68% (valence) on DREAMER.
  • The authors report MSC-TimesNet exceeded comparison models by no less than 23.24% on arousal and at least 21.89% on valence in the subject-independent setting, where MLP, logistic regression, and LSTM results were "all around 60%."
  • In the summary table, MSC-TimesNet outperformed EEGNet, MLP, LSTM, and logistic regression on both datasets and both experiment types; e.g., subject-independent DEAP arousal was 0.9723 for MSC-TimesNet versus 0.7399 for EEGNet.

Methods

  • Design: EEG is split into five frequency bands (δ, θ, α, β, γ), segmented into 0.5 s windows (6 baseline segments = 3 s; 120 emotion segments = 60 s), with PSD and DE features extracted per window; baseline feature means are subtracted from emotion features, and the resulting PSD+DE features are fused by a denoising autoencoder (DAE).
  • Model: MSC-TimesNet reshapes 1D features into 2D tensors using FFT-derived period information, applies dynamic convolutions along time (kernels 4, 8, 16) and channel (kernels 4, 2) axes, aggregates features by amplitude-weighted Softmax, and classifies via dropout + GeLU + linear layers.
  • Data and analysis: public DEAP (32 participants, 40 videos, 32 EEG channels, 512 Hz) and DREAMER (23 participants, 18 clips, 16 channels, 128 Hz) datasets; signals resampled to 128 Hz; labels thresholded at 5 (DEAP) and 3 (DREAMER) for high/low arousal and positive/negative valence; 5-fold cross-validation; Adam optimizer, learning rate 1e-4, weight decay 0.06, cross-entropy loss, 500 iterations; accuracy and F score as metrics.

Limitations

  • Stated: reliance on deep learning for automatic feature extraction may fail to capture rapidly changing, fine-grained time–frequency dynamics in non-stationary emotional signals, potentially losing critical information; the authors suggest synchrosqueezed and tunable-Q wavelet transforms as future remedies.
  • Stated: current feature extraction focuses on single-channel analysis, making it difficult to capture cross-channel synergistic properties and global dynamic patterns; multivariate time–frequency methods are proposed as future work.
  • Evident from the text: the reported subject-dependent accuracies (up to 98.97%) are far above the comparison methods and typical published DEAP results, and the paper offers no ablation isolating the contribution of the DAE fusion versus MSC-TimesNet; also, the dataset table lists DREAMER as having 14 EEG channels while the dataset description in the same section states 16 channels, an internal inconsistency.

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

Integrity signals

Registry

Crossref · 10.1080/10255842.2025.2477801

Cites

46 of 48 references verified0 citation errors · 2 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

3 Valid · 1 Not valid · 9 Not checked · 22 Own results and interpretations

4 more claims

Create a free account to see all 35 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 12
  • It may state two separate claims as one. 15
  • The quote only partly states this claim. 3

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.

2 of 2

Citations

Cites

Show all 42 references

Research works printed but not linked: 1

  1. Bellman R. 1957. Dynamic programming. Princeton, NJ: Princeton University Press; p. 4–9.

Cited by

See all 41

Metadata

Authors
Yin, J., Qiao, Z., Han, L., & Zhang, X.
Affiliations
  • Kunming University of Science and Technology (CN)
  • Zhejiang University (CN)
Corresponding author
Zhang, Xiangliang
Venue
Computer Methods in Biomechanics and Biomedical Engineering
Published
2025-03-17
Volume
29
Issue
10
Pages
2397-2414
ISSN
  • 1025-5842 (ISSN-L)
  • 1476-8259
OpenAlex
W4408518779
Cited by
41
References
42
Citations by year
2026: 13 · 2025: 28
FWCI
42.63
Type
journal article
Language
en
Open access
closed
Licence
doi.org/10.1080/policies-text-data-mining (tdm)
Locations
  • Computer Methods in Biomechanics & Biomedical Engineering (publishedVersion)
  • PubMed (publishedVersion)
Funders
National Natural Science Foundation of China (61741206)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Emotion and Mood Recognition
  • Life Sciences › Neuroscience › Cognitive Neuroscience › EEG and Brain-Computer Interfaces
  • Physical Sciences › Engineering › Biomedical Engineering › Muscle activation and electromyography studies
Keywords
EEG-based emotion recognition, TimesNet, multi-scale convolutional kernels, power spectral density features, differential entropy features, frequency band segmentation, DEAP dataset, DREAMER dataset
MeSH terms
Deep Learning, Autoencoder, Algorithms, Electroencephalography, Emotions, Humans, Signal Processing, Computer-Assisted

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 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

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: EEG-based emotion recognition with autoencoder feature fusion and MSC-TimesNet model. https://app.cochat.ai/papers/10.1080/10255842.2025.2477801
The paper
Yin, J., Qiao, Z., Han, L., & Zhang, X. (2026). EEG-based emotion recognition with autoencoder feature fusion and MSC-TimesNet model. Computer Methods in Biomechanics and Biomedical Engineering. https://doi.org/10.1080/10255842.2025.2477801

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.