ERTNet: an interpretable transformer-based framework for EEG emotion recognition.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal analysis is crucial for precise and immediate clinical assessment of emotional states.
• The complexity of EEG data challenges traditional recognition methods, while deep learning offers potential but often lacks interpretability.
• Accurate and interpretable emotion recognition from EEG is vital for advancing brain-computer interfaces.

Purpose of the Study:

• To develop an interpretable, end-to-end framework for emotion recognition using EEG signals.
• To leverage a hybrid Convolutional Neural Network (CNN) and Transformer architecture for enhanced spatiotemporal feature extraction.
• To improve the accuracy and interpretability of deep learning models in EEG-based emotion recognition.

Main Methods:

• A hybrid CNN-Transformer architecture was employed, integrating temporal and spatial convolutions.
• Temporal convolution focused on isolating salient EEG information and filtering noise.
• The Transformer module processed feature maps to capture high-level spatiotemporal features for emotion identification.

Main Results:

• The proposed model achieved high accuracy in diverse emotion classification tasks: 74.23% ± 2.59% on the DEAP dataset (dimensional) and 67.17% ± 1.70% on the SEED-V dataset (discrete).
• Performance surpassed existing CNN and Long Short-Term Memory (LSTM)-based models.
• Interpretive analysis revealed that beta and gamma EEG bands significantly influence emotion recognition, and the model effectively filters high-frequency noise.

Conclusions:

• The developed framework offers a promising, interpretable, and accurate solution for EEG-driven emotion recognition.
• Its ability to tailor convolution kernels enhances noise filtering and model robustness.
• This research paves the way for more sophisticated EEG-based emotion brain-computer interfaces.