A lightweight transformer model for robust EEG emotion recognition using channel-wise differential entropy

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Emotion recognition from electroencephalogram (EEG) signals is crucial for applications in healthcare, human-computer interaction, and education.
• Efficient and accurate decoding of emotional information from EEG is a significant research challenge.

Purpose of the Study:

• To propose a novel EEG emotion recognition model, the Channel-wise Differential Entropy Transformer (CWDET).
• To leverage differential entropy (DE) features and Transformer encoder for enhanced emotion recognition accuracy and efficiency.

Main Methods:

• DE features were extracted from EEG signals across five frequency bands (δ, θ, α, β, γ).
• EEG channels were processed as independent input tokens, mapped to high-dimensional space via embedding and positional encoding.
• A multi-head self-attention mechanism was employed for global feature fusion across channels, reducing redundancy and computational cost.

Main Results:

• The CWDET model achieved high classification accuracies of 98.63% on the SEED dataset and 99.16% on the SEED-IV dataset.
• The model demonstrated excellent stability and automatically focused on key brain regions.
• Even with a subset of channels, 93.44% recognition performance was maintained on the SEED-IV dataset.

Conclusions:

• CWDET offers a simple structure, computational efficiency, and high performance for EEG emotion recognition.
• The model provides a feasible low-resource solution for practical applications.
• This work supports the development of EEG emotion recognition technology and future generalization research.