A multi-band spatial asymmetry convolutional neural network for EEG-based emotion recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Electroencephalogram (EEG)-based emotion recognition is crucial for human-computer interaction.
• Existing methods often overlook spectral information and hemispheric asymmetry, limiting accuracy.

Purpose of the Study:

• To propose a novel Multi-Band Spatial Asymmetry Convolutional Neural Network (MBSACNN) for enhanced EEG-based emotion recognition.
• To capture complementary spectral information and left-right hemispheric asymmetry in EEG signals.

Main Methods:

• EEG signals decomposed into theta, alpha, beta, and gamma bands.
• Constructed dual-input representations: Original EEG Matrix (OEM) and Spatial Asymmetric EEG Matrix (SAEM).
• Utilized a 2D Convolutional Neural Network (CNN) for feature extraction.

Main Results:

• Achieved high average accuracies (97.07% arousal, 96.61% valence) and F1-scores (97.19% arousal, 96.89% valence) on the DEAP dataset.
• Demonstrated superior performance compared to multiple conventional and deep-learning baselines.
• Ablation studies confirmed the contributions of multi-band decomposition, spatial asymmetry, and specific CNN parameters.

Conclusions:

• MBSACNN effectively integrates multi-band spectral and spatial asymmetry information for robust emotion recognition.
• The proposed model offers a significant advancement in EEG-based emotion recognition accuracy and stability.
• This approach holds promise for developing more sophisticated human-computer interaction systems.