MGSN: Depression EEG lightweight detection based on multiscale DGCN and SNN for multichannel topology

As the global public health risk intensifies, the number of patients with depression is increasing. Since the current clinical scale assessments may be influenced by subjective patient factors and physician diagnostic experience, we need to explore objective biomarkers from complex electroencephalographic (EEG). Addressing the multichannel topology characteristic of depression, we innovatively propose a lightweight depression detection method based on multiscale dynamic graph convolutional networks and spiking neural networks. By constructing a foundational detection framework utilizing a spiking neural network, the model processes information in the form of discrete spikes and highly fits biological neuron mechanisms. To handle the complexity issue arising from topological information and channel features, we introduce dynamic graph convolution for effective spatiotemporal attribute aggregation. Moreover, to circumvent the costly resource consumption associated with graph computation, we design multiple diffusion branches with different receptive field levels, and obtain multiscale topological information in parallel. By strengthening the learning of neighboring node information, the framework is optimized. Additionally, the integration of EEG under both positive and negative stimulation significantly improves the model's representation of multichannel topology. Our method achieves a classification accuracy of 90.51%, and improves the detection efficiency without neglecting multichannel structural relationships. In addition, after visualizing the network output features, it shows that patients with depression exhibit distinct frontal and temporal EEG abnormalities compared to healthy controls.