EEG Emotion Recognition via Graph-based Spatio-Temporal Attention Neural Networks

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Emotion recognition from electroencephalography (EEG) signals is crucial for affective computing and brain-computer interfaces (BCI).
• Extracting discriminative features from complex EEG data remains a significant challenge for existing deep learning models.
• Current methods often struggle to effectively capture both spatial and temporal dynamics inherent in EEG signals.

Purpose of the Study:

• To propose a novel Spatio-Temporal Attention Neural Network (STANN) for enhanced EEG-based emotion recognition.
• To effectively extract discriminative spatial and temporal features from EEG signals.
• To explore the utility of graph signal processing (GSP) for representing inter-channel EEG relationships.

Main Methods:

• Developed a Spatio-Temporal Attention Neural Network (STANN) utilizing a parallel structure of multi-column convolutional neural networks and attention-based bidirectional long-short term memory.
• Integrated graph signal processing (GSP) tools to analyze and incorporate inter-channel EEG signal relationships, forming the GFT-STANN architecture.
• Conducted subject-wise classification experiments for binary (valence, arousal) and four-class emotion recognition tasks.

Main Results:

• The proposed STANN model significantly improved the state-of-the-art performance in subject-wise emotion recognition tasks.
• Both raw EEG signals and their graph representations (GFT-STANN) input to the network yielded superior classification accuracy.
• The network effectively extracted discriminative spatio-temporal features, outperforming existing methods.

Conclusions:

• The novel STANN architecture offers a powerful approach for accurate emotion recognition from EEG signals.
• Incorporating GSP enhances the model's ability to leverage inter-channel information, further boosting performance.
• This work advances the capabilities of affective computing and BCI systems through improved EEG signal analysis.