Related Experiment Video
Updated: Jan 9, 2026
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
5.2K
Traditional Machine Learning Outperforms EEGNet for Consumer-Grade EEG Emotion Recognition: A Comprehensive
Carlos Rodrigo Paredes Ocaranza1, Bensheng Yun1, Enrique Daniel Paredes Ocaranza1
1School of Artificial Intelligence and Information Engineering, Zhejiang University of Science and Technology, Hangzhou 310023, China.
Sensors (Basel, Switzerland)
|December 11, 2025
Summary
Traditional machine learning (ML) with feature engineering significantly outperformed complex deep learning models like EEGNet for consumer-grade emotion recognition. This approach offers superior accuracy, stability, and efficiency in noisy, real-world brain-computer interface applications.
Area of Science:
- Neuroscience
- Computer Science
- Biomedical Engineering
Background:
- Consumer-grade EEG devices offer potential for widespread brain-computer interface (BCI) deployment.
- Challenges include reduced spatial resolution, variable signal quality in uncontrolled environments, and limited exploration of deep learning efficacy and generalizability.
Purpose of the Study:
- To systematically compare traditional machine learning (ML) with domain-specific feature engineering against the EEGNet deep learning architecture for emotion recognition using consumer-grade EEG.
- To determine when and why feature engineering outperforms end-to-end learning in resource-constrained scenarios.
Main Methods:
- Within-dataset evaluation using the DREAMER dataset and cross-dataset validation (DREAMER→SEED-VII).
- Traditional ML utilized statistical, frequency-domain, and connectivity features with random forest classification.
- Deep learning employed optimized EEGNet architectures; cross-dataset validation used progressive domain adaptation (anatomical mapping, CORAL, TCA).
Main Results:
- Traditional ML achieved superior within-dataset performance (F1=0.945 vs. 0.567, p<0.000001) and cross-dataset performance (F1=0.619 vs. 0.007).
- Inter-channel connectivity features contributed 61% discriminative power; traditional ML showed 95% faster training and 10x faster inference.
- Deep learning degraded more under noise (17% vs. <1%), while traditional ML remained robust.
Conclusions:
- Architectural complexity does not universally enhance biosignal processing for consumer-grade EEG.
- Domain-specific feature engineering and lightweight adaptation offer superior accuracy, stability, and practical deployment for EEG emotion recognition.
- Findings suggest principles extend to other complex architectures, enabling robust cross-system BCI applications.
