Cerebral asymmetry representation learning-based deep subdomain adaptation network for electroencephalogram-based...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Signal Processing

Background:

• Extracting spatial information from electroencephalogram (EEG) for emotion recognition is challenging.
• Individual differences cause domain shifts, degrading cross-subject EEG classification performance.

Purpose of the Study:

• To propose the cerebral asymmetry representation learning-based deep subdomain adaptation network (CARL-DSAN) for enhanced cross-subject EEG-based emotion recognition.
• To address challenges in spatial information extraction and domain shift in EEG emotion recognition.

Main Methods:

• The CARL module uses a two-step strategy for intra-hemisphere spatial learning and asymmetry representation learning, inspired by brain hemisphere activity.
• Transformer encoders in CARL emphasize contributive electrodes and pairs.
• The DSAN module mitigates domain shift by aligning relevant subdomains for improved cross-subject performance.

Main Results:

• Subject-independent experiments on the DEAP database yielded accuracies of 68.67% for arousal and 67.11% for valence.
• On the MAHNOB-HCI database, accuracies were 67.70% for arousal and 67.18% for valence.

Conclusions:

• The CARL-DSAN model demonstrates outstanding cross-subject performance in both arousal and valence classification.
• The proposed method effectively enhances EEG-based emotion recognition by leveraging cerebral asymmetry and subdomain adaptation.