A Saliency based Feature Fusion Model for EEG Emotion Estimation.

Area of Science:

• Neuroscience
• Computer Science
• Affective Computing

Background:

• Electroencephalogram (EEG) is a key modality for assessing brain activity and has shown promise in emotion estimation.
• EEG-based emotion recognition has potential applications in diagnosing and rehabilitating neurological and psychological conditions.
• Current methods for EEG emotion analysis can be limited in accuracy and stability.

Purpose of the Study:

• To propose a novel dual model for enhanced emotion estimation from EEG data.
• To explore two distinct representations of EEG feature maps: sequential band power and image-based feature vectors.
• To introduce an innovative saliency-based method for combining these representations to promote joint learning.

Main Methods:

• A dual model approach was developed, processing EEG data through both sequential and image-based feature representations.
• EEG band power was utilized for the sequential representation, while feature vectors formed the image-based representation.
• A saliency analysis method was employed to integrate information from both representations, facilitating joint model learning.

Main Results:

• The proposed dual model achieved state-of-the-art performance on three out of four publicly available datasets (SEED-IV, SEED, DEAP, MPED).
• The model demonstrated improved stability, indicated by a lower standard deviation in results compared to existing approaches.
• The joint learning strategy effectively combined information from both sequential and image-based EEG representations.

Conclusions:

• The developed dual model offers a more accurate and stable approach to EEG-based emotion estimation.
• The integration of sequential and image-based EEG features via saliency analysis is a promising direction for affective computing.
• The findings suggest significant potential for EEG in clinical applications related to emotion assessment and mental health.