A Novel EEG-Based Hypergraph Convolution Network for Depression Detection: Incorporating Unified Brain Network and Multi-Segment Spatiotemporal EEG Features

Detecting depression from Electroencephalography (EEG) signals remains a challenging task due to the complexity of brain networks and the significant individual differences in neural activity. Traditional models significantly fall short: 1) capturing the EEG brain connection beyond just pairwise interactions, 2) Ignores the inter-channel spatial relations as it plays a critical role in aggregating features from other channels and 3) While existing studies have made significant strides in capturing temporal dependencies within brain networks, they lack a comprehensive method for effectively fusing these dependencies. Thus, to address these issues, the paper proposes hierarchical fusion results in a unified Subject-Subject Brain Network, capturing spatio-temporal multi-segment EEG features and channel-channel local relationships at a low-level using Graph Convolution Gated Recurrent unit and integrating hypergraph-based subject-specific data at a higher-level using Hypergraph Convolution Network (HGCN) for depression detection. The model shows high independence among the EEG features and achieves 83% accuracy for Major depressive disorder (MDD). The suggested model exhibits a 16.87% enhancement in accuracy compared to SOTA models. This indicates that the model effectively retains the independence of spatial and temporal MDD features from normal features, highlighting its potential for accurate real-time applications.