Decoding Non-Invasive Electroencephalography Signal via a Two-Discriminator Adversarial Network.

Area of Science:

• Neuroscience and Artificial Intelligence
• Biosignal Processing and Machine Learning

Background:

• Electroencephalography (EEG) records brain activity for emotion decoding but faces challenges due to individual brain differences and complex channel interrelationships.
• Existing non-invasive biosensing methods struggle with accuracy in emotion recognition because of inter-subject variability and intricate EEG signal patterns.

Purpose of the Study:

• To propose a Two-Discriminator Domain Adversarial Neural Network (TD-DANN) for enhanced emotion decoding from EEG signals.
• To achieve more generalized and individualized emotion feature representations through adversarial learning for improved accuracy in non-invasive biosensing.

Main Methods:

• Utilized graph convolution to extract features from EEG signals, modeling channels as graph nodes with a dynamically learned adjacency matrix.
• Implemented a two-discriminator approach: a domain discriminator for universal features and an individual discriminator for personalized emotion adaptability.
• Employed adversarial learning to minimize feature distribution differences between source and target domains, enhancing feature universality and individual consistency.

Main Results:

• Achieved high subject-dependent recognition accuracies of (98.45 ± 2.38)% on the SEED dataset and (84.40 ± 8.70)% on the SEED-IV dataset.
• Demonstrated strong subject-independent recognition accuracies of (89.45 ± 5.87)% on the SEED dataset and (77.13 ± 7.97)% on the SEED-IV dataset.
• The TD-DANN method significantly improved emotion decoding accuracy in both subject-dependent and subject-independent scenarios across different datasets.

Conclusions:

• The proposed TD-DANN effectively addresses the challenges of individual differences and complex channel interrelationships in EEG-based emotion decoding.
• The method's ability to learn generalized and individualized features validates its effectiveness for accurate and adaptive emotion recognition in non-invasive biosensing.
• The experimental results confirm the TD-DANN's superior performance, highlighting its potential for practical applications in affective computing and brain-computer interfaces.