FetchEEG: a hybrid approach combining feature extraction and temporal-channel joint attention for EEG-based emotion...

Area of Science:

• Neural Engineering
• Brain-Computer Interfaces
• Affective Computing

Background:

• Electroencephalogram (EEG) analysis is crucial for neural engineering tasks like emotion recognition.
• Traditional feature extraction methods show promise, but deep learning's end-to-end approaches often neglect channel representations and face model fitting challenges.

Purpose of the Study:

• To introduce FetchEEG, a hybrid approach for emotion classification using EEG data.
• To address limitations of current deep learning methods by integrating feature extraction with temporal-channel joint attention.

Main Methods:

• FetchEEG employs a multi-head self-attention mechanism for simultaneous extraction of temporal and channel representations from EEG data.
• Joint representations are concatenated and classified using fully-connected layers.
• Performance is validated on self-developed and public EEG datasets.

Main Results:

• FetchEEG outperforms state-of-the-art methods in both subject-dependent and independent emotion recognition tasks across all tested datasets.
• The study analyzes the impact of sliding window parameters and frequency bands on recognition sensitivity.
• FetchEEG demonstrates stronger generalization ability compared to existing methods.

Conclusions:

• FetchEEG offers a novel, effective, and feasible hybrid method for EEG-based emotion classification.
• It achieves state-of-the-art results with significantly improved training efficiency over end-to-end deep learning models.