Domain-generalized Deep Learning for Improved Subject-independent Emotion Recognition Based on Electroencephalography

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) offers noninvasive, high temporal resolution for applications like emotion recognition.
• Subject variability in EEG data presents a major challenge for developing generalizable emotion recognition models.
• Existing deep learning models struggle with intersubject and intersession variations, limiting their real-world applicability.

Purpose of the Study:

• To systematically evaluate domain generalization (DG) techniques for subject-independent EEG-based emotion recognition.
• To investigate the effectiveness of combining four DG methods (Deep CORAL, GroupDRO, VREx, DANN) with three deep learning architectures (ShallowFBCSPNet, EEGNet, TSception).
• To assess the performance of these integrated models on two distinct emotional EEG datasets.

Main Methods:

• Twelve distinct models were created by combining four DG techniques with three deep learning architectures.
• Each subject's data was treated as a separate domain to simulate real-world variability.
• Binary classification tasks for valence and arousal states were performed using a ten-fold cross-validation strategy.

Main Results:

• Domain generalization methods consistently improved classification accuracy across both datasets.
• The TSception architecture combined with VREx achieved top performance for both valence and arousal in one dataset.
• In the second dataset, TSception with VREx led in valence classification, while TSception with GroupDRO excelled in arousal classification.

Conclusions:

• Domain generalization approaches effectively mitigate distributional shifts caused by intersubject and intersession variability in EEG data.
• The integration of DG techniques with deep learning architectures holds significant promise for developing robust, subject-independent EEG-based emotion recognition systems.
• Specific DG methods like VREx and GroupDRO, particularly with the TSception architecture, demonstrate superior performance in enhancing emotion recognition accuracy.