A Depression Level Classification Model Based on EEG and GCN Network with Domain Generalization

Electroencephalography (EEG) contains a wealth of information, including neuron activity, allowing for a partial understanding of the brain’s state. As a reliable tool, EEG, combined with deep neural networks, has been increasingly utilized to assist in the detection of depression and has achieved satisfactory results. However, there are still some limitations in the current models. Firstly, it is worth noting that many existing models are binary classification models. These models can determine whether a subject is depressed or not, but they lack the capability to differentiate the severity of depression. Besides, the variations in EEG signals among various individuals have not been thoroughly taken into account, leading to subpar performance when applying the trained model to new subjects. To address the aforementioned issues, we suggest implementing a multi-classification model that utilizes a graph convolutional neural network (GCN) to identify levels of depression. Additionally, to mitigate the problem of imbalanced samples in multi-classification tasks, we introduce a penalty coefficient for the smaller categories. At the same time, the issue of signal discrepancies among subjects was taken into account, leading to the introduction of the domain generalization module. Finally, the depression level identification task on the MODMA and PRED+CT dataset achieved an accuracy of $75.47 \%$ and $77.97 \%$ respectively, surpassing the performance of the state-of-the-art (SOTA) model. In addition, numerous ablation experiments were conducted to validate the efficacy of each module.