Cross-subject EEG Channel Optimization by Domain Adversarial Sparse Learning Model

How to decrease the number of electroencephalogram (EEG) record channels, and acquire the optimal electrodes to perform EEG signals analysis, are of extremely importance in developing and promoting highly available Brain-Computer Interface (BCI). In this paper, we design an EEG channel optimization model, named Domain Adversarial Sparse Learning model (DASL), to perform fatigue state detection with minimal and optimal EEG electrodes. DASL composes of Sparse Learning (SL), Domain Adversarial Neural Networks (DANN) and Generative Adversarial Networks (GAN). Herein, SL is used to find the optimal EEG channels through selecting key features from the source domain, these key features are then used to determine fatigue state by DANN across subjects, GAN aims at improving the robustness for our proposed model. Experimental results show DASL outperforms other traditional machine learning methods in the classification performance of mental state tasks under the condition of optimal and minimal EEG electrodes.