Symmetric Convolutional and Adversarial Neural Network Enables Improved Mental Stress Classification From EEG

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for mental stress classification.
• EEG signal variability presents challenges in feature extraction and cross-subject generalization.
• Existing methods struggle with robust and adaptable stress detection from EEG.

Purpose of the Study:

• To propose a novel deep neural network, Symmetric Deep Convolutional Adversarial Network (SDCAN), for enhanced mental stress classification using EEG.
• To improve classification accuracy and generalization across subjects by leveraging adversarial inference for invariant feature extraction.
• To validate the effectiveness of SDCAN compared to conventional methods in classifying stress stages.

Main Methods:

• Developed a Symmetric Deep Convolutional Adversarial Network (SDCAN) integrating convolutional neural networks (CNNs) and adversarial theory.
• Utilized adversarial inference to automatically extract invariant and discriminative features from raw EEG data.
• Conducted experiments with 22 subjects undergoing the Trier Social Stress Test, collecting EEG data and calibrating stress into four or five stages based on salivary cortisol levels.
• Applied Euclidean space data alignment (EA) and leave-one-subject-out cross-validation to assess cross-subject generalization.

Main Results:

• SDCAN achieved classification accuracies of 87.62% for four stages and 81.45% for five stages, outperforming conventional CNN methods.
• The EA-SDCAN model demonstrated improved cross-subject generalization, achieving accuracies of 60.52% (four stages) and 48.17% (five stages) via leave-one-subject-out cross-validation.
• Adversarial inference successfully captured invariant and discriminative EEG features for stress classification.

Conclusions:

• The proposed SDCAN network offers a more feasible and effective approach for classifying mental stress stages from EEG data.
• SDCAN enhances classification accuracy and generalization capabilities, addressing key limitations of current EEG-based stress detection methods.
• The integration of adversarial theory provides a promising direction for developing robust brain-computer interfaces for mental state monitoring.