Cross-Layer Residual-Convolutional Fusion Network for Tri-Class Depression Diagnosis from EEG Signals

Depression profoundly impacts human lives and casts a significant shadow over societal well-being. Electroencephalography (EEG) has emerged as a crucial tool in assisting with the treatment of this pervasive illness. While traditional EEG-Conformer frameworks successfully integrate the strengths of Convolutional Neural Networks (CNNs) and Transformers, their initial shallow convolutional layers may sometimes be insufficient to capture the intricate local features within EEG signals. This study introduces ConvResFormer, a novel deep learning model for EEG-based depression detection. ConvResFormer incorporates two key enhancements:An additional convolutional layer: This layer supports enhanced local feature extraction, improving the model's ability to discern subtle patterns within EEG data.A residual network after the Transformer encoder: This addition bolsters the model's robustness and generalization capabilities. Our research demonstrates that ConvResFormer achieves an average classification accuracy of 76.67 %, significantly outperforming traditional models and offering a promising avenue for improving depression diagnosis and treatment.