EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Robustness of Lightweight Neural Networks for EEG-Based Emotion Recognition

Sivaraj Nimishan, Selvarajah Thuseethan, Shanmuganathan Vasanthapriyan, Roshan G. Ragel, Romiyal George

IEEE Access · 2026 · 14 · 96497-96511

DOI: 10.1109/access.2026.3706425 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Sivaraj Nimishan und 4 weiteren Autor:innen. Die Arbeit nennt 52 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: LK Domain Registry for Open Access Publishing.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Electroencephalography (EEG) is widely regarded as a suitable modality for emotion recognition, as it records neural signals that are not easily influenced by conscious behaviour. However, real-world deployment of such systems is limited by the complexity of deep learning (DL) models. Existing EEG-based studies primarily focus on improving the accuracy of DL models but often overlook their robustness and efficiency in resource-constrained and noisy settings. In this study, we address this gap by systematically evaluating the feasibility of lightweight DL models for EEG-based emotion recognition. We propose a two-stage pipeline that combines wavelet-based artefact removal with classification. The framework uses various widely used DL architectures, such as MLP, LSTM, GRU, alongside newly adapted 1D variants of ResNet, VGG, MobileNet, and SqueezeNet, operating directly on raw EEG signals, eliminating the need for image-based transformations. Extensive experiments are conducted on four benchmark EEG datasets, such as DEAP, DREAMER, LUMED and GAMEEMO, using evaluation metrics such as accuracy, F1 score, model parameters, training time, inference time, and model size. Among all models, ResNet1D achieved the highest classification accuracy and F1 score of 92.85% and 92.67%, with only 473,865 parameters, outperforming VGG1D, MobileNet1D, and SqueezeNet1D by 1.74%, 8.89%, and 7.62%, respectively. Despite their lower accuracy than ResNet1D, MobileNet1D and SqueezeNet1D demonstrated superior efficiency, with model sizes as low as 0.215 MB and 2.09 MB, respectively, and inference times under 0.11 seconds. Overall, the findings indicate that lightweight ResNet1D models strike an effective balance between accuracy and efficiency.

Förderung

  • LK Domain Registry for Open Access Publishing

Zitieren

APA

Nimishan, S., Thuseethan, S., Vasanthapriyan, S., Ragel, R. G., & George, R. (2026). Robustness of Lightweight Neural Networks for EEG-Based Emotion Recognition. IEEE Access, 14, 96497-96511. https://doi.org/10.1109/access.2026.3706425

Vancouver

Nimishan S, Thuseethan S, Vasanthapriyan S, Ragel RG, George R. Robustness of Lightweight Neural Networks for EEG-Based Emotion Recognition. IEEE Access. 2026;14:96497-96511. doi:10.1109/access.2026.3706425

BibTeX
@article{Nimishan2026,
  author = {Nimishan, Sivaraj and Thuseethan, Selvarajah and Vasanthapriyan, Shanmuganathan and Ragel, Roshan G. and George, Romiyal},
  title = {Robustness of Lightweight Neural Networks for EEG-Based Emotion Recognition},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {96497--96511},
  doi = {10.1109/access.2026.3706425}
}
Kostenlos starten →