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

With the increasing demand for emotion recognition technology in fields such as healthcare, human-computer interaction, and education, the efficient and accurate decoding of emotional information from electroencephalogram (EEG) signals has become a research hotspot. This study proposes a brain EEG emotion recognition model, the Channel-wise Differential Entropy Transformer (CWDET), based on the combination of differential entropy (DE) features and a transformer encoder. In this method, the DE features of the EEG signals are first extracted in five frequency bands:δ,θ,α,β, andγ. Each channel is treated as an independent input token, and through simple but efficient embedding and positional encoding, low-dimensional information is mapped into a high-dimensional space. A multi-head self-attention mechanism is then employed to achieve global feature fusion across channels, effectively reducing data redundancy and computational cost. The experiments conducted on the SEED and SEED-IV datasets achieved high classification accuracies of 98.63% and 99.16%, respectively, with the model performing excellently in terms of the standard deviation and stability. Further analysis of the attention weights revealed that the model automatically focused on key brain regions, such as the prefrontal area, central, and centro-parietal junction. Even when only a subset of channels was selected, the model maintained a recognition performance of 93.44% on the SEED-IV dataset. Comparative experiments with various existing advanced methods show that the CWDET offers a simple structure and computational efficiency while maintaining high performance, providing a feasible low-resource solution for practical EEG emotion recognition applications. This study not only provides new theoretical and practical support for the development of EEG emotion recognition technology but also lays a solid foundation for future generalization research across subjects and sessions.