CDBA: a novel multi-branch feature fusion model for EEG-based emotion recognition

Area of Science:

• Biomedical engineering
• Machine learning
• Artificial intelligence

Background:

• Traditional Electroencephalography (EEG)-based emotion recognition faces limitations due to single-feature input and the nonlinear nature of EEG signals, hindering real-time brain-computer interface applications.
• Existing time and frequency domain methods are insufficient for capturing complex, nonlinear EEG signal characteristics.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient EEG-based emotion recognition.
• To address the limitations of traditional methods by incorporating multi-branch feature extraction and an attention mechanism.

Main Methods:

• A novel CNN-DSC-Bi-LSTM-Attention (CDBA) model was proposed, utilizing multi-branch feature extraction from normalized EEG signals.
• Features were concatenated, and an attention mechanism layer assigned weights to each channel's feature.
• Softmax classification was employed for EEG signal categorization.

Main Results:

• The CDBA model demonstrated high classification accuracies: 99.44% for triple-category and 99.99% for four-category emotions on the SEED dataset.
• On the DREAMER dataset, the model achieved 84.49% accuracy for five-category valence classification.
• Multi-classification experiments using ten-fold cross-validation showed superior performance and generalization compared to other models.

Conclusions:

• The proposed multi-branch feature fusion deep learning model with an attention mechanism effectively handles nonlinear EEG data for emotion recognition.
• The CDBA model exhibits strong fitting and generalization capabilities, proving its efficacy for advanced brain-computer interface systems.
• This approach holds significant potential for diagnosing and treating nervous system diseases and enhancing emotion-based BCI applications.