SEEG Emotion Recognition Based on Transformer Network With Channel Selection and Explainability

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Brain-computer interface (BCI) technology is crucial for treating refractory emotional disorders.
• Stereo-electroencephalography (SEEG) offers precise neural activity recording from deep brain structures and the cortex.
• SEEG has significant potential for developing emotion recognition BCIs.

Purpose of the Study:

• To develop and evaluate a novel algorithm for emotion recognition using SEEG data.
• To construct an emotion dataset from SEEG recordings of nine subjects.
• To assess the performance of the proposed Spatial Transformer-based Hybrid Network (STHN) against existing methods.

Main Methods:

• Collected SEEG data from nine subjects to create an emotion dataset.
• Developed a Spatial Transformer-based Hybrid Network (STHN) for SEEG emotion recognition.
• Evaluated STHN's performance against baseline methods like EEGNet, TSception, and deep convolution neural networks.

Main Results:

• STHN achieved a triple-classification accuracy of 83.56%, outperforming baseline methods.
• STHN demonstrated channel weighting and selection capabilities, identifying key brain regions for emotion recognition.
• High-weight channels were predominantly in emotion-associated areas like the frontal lobe, temporal lobe, and hippocampus, indicating model explainability.

Conclusions:

• The developed SEEG emotion recognition algorithm using STHN is effective and offers a degree of explainability.
• This novel approach holds significant potential for monitoring and treating patients with refractory emotional disorders.
• This study represents the first development of an SEEG-based emotion recognition algorithm.