A novel EEG-based graph convolution network for depression detection: Incorporating secondary subject partitioning and attention mechanism.

Electroencephalography (EEG) is capable of capturing the evocative neural information within the brain. As a result, it has been increasingly used for identifying neurological disorders, such as depression. In recent years, researchers have proposed deep-learning models for EEG-based depression detection and achieved good results. However, there are still some limitations in these models, as the varying importance across different EEG channels and the varying importance of different features within the same channel for each subject have not been adequately addressed. Furthermore, the variations in EEG data distributions among different subjects have not been fully considered, thereby compromising the universality of the model in cross-subject tasks. To address the aforementioned problems, we propose a model with a secondary subject partitioning and attention mechanism based on a graph convolution network (GCN). First, we present an attention module that can simultaneously concentrate on multiple channels with different features within each channel. Second, domain generalization based on adversarial training is added to the model, and a secondary subject partitioning method is proposed to group subjects with similar data distributions into the same domain with a shared domain label. This effectively reduces the number of domain labels and increases the data volume in each domain, thereby enhancing the domain generalization performance. Finally, in the depression recognition task, the improved domain generalization and attention modules collaborate to capture subject-invariant features. Prediction accuracies of 92.87% and 83.17% are respectively achieved on two public datasets, outperforming the state-of-the-art baseline models. Moreover, extensive ablation experiments further validate the effectiveness of each module in the model. [ABSTRACT FROM AUTHOR]