LightK-DSGCN: Depression Detection in EEGs with Lightweight Kalman Filter-aided Dual-Stream Graph Convolutional Networks

Automated detection of depression using Electroencephalogram (EEG) signals is crucial for advanced disease treatment. However, existing EEG detection models still face challenges. 1) EEG signals are susceptible to noise interference and exhibit a high level of randomness. 2) Manual denoising and feature selection potentially introduce human bias. 3) The integrated message propagation across both spatial and temporal domains is not fully explored. Therefore, this paper proposes LightK-DSGCN, Enhancing Depression Detection with Lightweight Kalman Filter-aided Dual-Stream Graph Convolutional Networks, a novel framework for identifying characteristics EEG patterns of depression patients. LightK-DSGCN leverages dual-stream graph neural networks to simultaneously explore spatiotemporal features, effectively capturing the distinctive patterns exhibited by depression patients. Firstly, the EEG signals are decomposed into temporal and spatial components at each time point. Then, the temporal features are embedded using a dilation temporal convolutional network, while the spatial features are obtained through a graph convolutional network. Moreover, a lightweight Kalman filter combined with recurrent neural networks is proposed to denoise and align the spatiotemporal features, enabling the extraction of detailed information from multiple perspectives. Experimental results on two real-world datasets demonstrate the superiority of our LightK-DSGCN over state-of-the-art methods in detecting depression using EEG signals. LightK-DSGCN provides a promising approach for automated depression detection in clinical practice. The code can be found here.