Feature Extraction With Stacked Autoencoders for EEG Channel Reduction in Emotion Recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Emotion recognition using electroencephalogram (EEG) signals is challenging due to complex feature extraction and the need for numerous channels.
• Existing methods require significant EEG channels, limiting practical applications and device miniaturization.

Purpose of the Study:

• To investigate the use of deep learning for reducing EEG channels in emotion recognition.
• To develop a feature analysis method and algorithm for optimizing EEG channel selection.
• To maintain the quality of EEG signals during channel reduction for accurate emotion classification.

Main Methods:

• Utilized a stacked autoencoder (SAE) network for optimal feature extraction from EEG signals.
• Employed SAEs to capture both linear and non-linear features representative of the EEG signal.
• Applied a support vector machine (SVM) classifier to evaluate the extracted features for emotion recognition.

Main Results:

• Achieved an accuracy of 75.7% for valence and 74.4% for arousal dimensions using SAE-extracted features.
• Demonstrated a significant reduction in EEG channels from 32 to 12.
• Identified distinct channel compositions for valence and arousal dimension detection.

Conclusions:

• Deep learning, specifically SAEs, can effectively reduce EEG channels for emotion recognition without compromising signal quality.
• The optimized feature extraction method enables the design of smaller, more practical EEG devices.
• The findings offer a pathway for more efficient and accessible brain-computer interfaces.