EEG Signal Analysis And Application Based On Adaptive Graph Representation

The exploration of the human brain has never stopped in the history of science.As a safe and economical electrophysiological monitoring method,EEG can reflect neural activities in the brain,and therefore assist brain research and relevant applications.It is hard to effectively extract information from complex EEG signals,which is the difficulty faced in the process of EEG application.With its powerful learning ability,deep learning methods have been applied to EEG classification tasks.The graph representation-based EEG classification converts EEG signals into graph representations described by node feature matrix and adjacency matrix,which can better capture the connections between different brain regions.The construction of high-quality brain graph representation is worth exploring since it is beneficial to the improvement of task performance.In this paper,we propose an adaptive graph representation construction method based on EEG signals,which automatically learns graph representation during training.The proposed method is applied in mental disease diagnosis and emotion recognition tasks.Details are as follows:(1)An adaptive node feature extraction method is proposed,which combines important information extracted from the frequency,spatial,and subspace domain.Features are automatically extracted from the frequency domain and mapped to multiple subspaces to improve the representation ability.After that,a three-dimensional attention mechanism is applied to make the model concentrate on important features from frequency,spatial and subspace domain.(2)An adaptive adjacency matrix construction method is proposed,which considers both subject-dependent and subject-independent information.The subject-dependent adjacency matrix captures the individual differences among subjects during training,and the subjectindependent adjacency matrix is responsible for obtaining the general information shared among all subjects.The above methods were applied to the task of the diagnose of Premenstrual syndrome,a mental disease with high incidence but hard to diagnose.It achieved 76.15% and 72.53%accuracy on cross-subject diagnostic tasks,respectively.Combining the two methods,a complete adaptive graph representation construction method is obtained,which achieves a cross-subject accuracy of 83.57% on the public emotion recognition dataset SEED.