EEG-based emotion classification using LSTM under new paradigm.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning is increasingly used for complex classification tasks, including emotion recognition from physiological signals like electroencephalogram (EEG).
• Existing research often focuses on spatial features of EEG, potentially overlooking temporal dynamics crucial for emotion classification.

Purpose of the Study:

• To model emotion classification from EEG as a sequence classification problem, capturing temporal features.
• To investigate the effectiveness of Long Short-Term Memory (LSTM) networks for emotion classification using EEG power band frequency features.
• To evaluate the impact of segment size on classification accuracy.

Main Methods:

• EEG data from 32 channels were segmented, and power band frequency features were extracted for each segment.
• Three Long Short-Term Memory (LSTM) models (LSTM1, LSTM2, LSTM3) with varying memory cells (32, 64, 128) were trained.
• Four-class emotion classification (HVHA, HVLA, LVHA, LVLA) was performed based on valence and arousal models.

Main Results:

• The LSTM3 model (128 cells) achieved the highest classification accuracy of 90%.
• LSTM1 (32 cells) and LSTM2 (64 cells) achieved accuracies of 85% and 89%, respectively.
• Smaller segment sizes generally led to higher classification accuracy with LSTM models.

Conclusions:

• LSTM networks are effective for emotion classification from EEG, particularly when modeling temporal sequences.
• The temporal dynamics of EEG signals contain significant information for distinguishing emotional states.
• Optimizing segment size is crucial for maximizing classification accuracy in LSTM-based EEG emotion recognition.