Simultaneously exploring multi-scale and asymmetric EEG features for emotion recognition.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Emotion recognition using electroencephalography (EEG) is a growing area in brain-computer interaction (BCI).
• Neuroscience indicates hemispheric asymmetry in brain activity during emotional states, a key principle for emotion recognition models.
• The non-stationary nature of EEG signals necessitates multi-scale feature extraction for accurate classification.

Purpose of the Study:

• To propose a novel deep learning model, the Multi-Scales Bi-hemispheric Asymmetric Model (MSBAM), for enhanced EEG-based emotion recognition.
• To leverage both multi-scale EEG signal characteristics and hemispheric asymmetry for improved classification accuracy.
• To validate the model's performance on established emotion datasets.

Main Methods:

• Developed a convolutional neural network (CNN) based model named MSBAM.
• Incorporated multi-scale convolution kernels to capture diverse EEG signal features.
• Exploited bi-hemispheric asymmetry in neural activity for model design.
• Evaluated the model on the public DEAP and DREAMER EEG datasets.

Main Results:

• MSBAM achieved over 99% accuracy in two-class emotion classification.
• The model successfully classified low-level and high-level states across arousal, valence, dominance, and liking dimensions.
• Demonstrated the effectiveness of multi-scale feature extraction and hemispheric asymmetry in EEG-based emotion recognition.

Conclusions:

• The MSBAM model shows significant promise for advancing EEG-based emotion recognition in BCI.
• The study highlights the importance of considering multi-scale signal properties and neural mechanisms for designing effective deep learning models.
• This approach offers a robust method for understanding and classifying human emotional states from neural data.