A multiscale spatio-temporal convolutional attention network for depressive disorder diagnosis via coarse-grained...

Depression disorder (DD) is a common mental disorder and a leading contributor to the global burden of disease. However, accurate diagnosis remains challenging due to the reliance of conventional approaches on subjective assessments. To address this limitation, this study proposes a Multiscale Spatio-temporal Convolutional Attention Network (MSTANet) for EEG-based depression detection. The proposed framework integrates multiscale coarse-grained signal reconstruction with localized spatio-temporal attention modeling, where the spatial dimension explicitly captures inter-channel relationships between bilateral frontal EEG signals (Fp1 and Fp2). MSTANet adopts a multi-parallel convolutional architecture combined with Squeeze-and-Excitation (SE) modules and a fully convolutional network (FCN), enabling effective multiscale spatio-temporal feature learning while maintaining a simplified dual-channel frontal EEG acquisition scheme for improved clinical practicality. Ablation and comparative experiments demonstrate that the proposed model outperforms existing methods, achieving an accuracy of 93.61%. To further assess robustness and clinical generalizability, a dual-center validation scheme was conducted. The model was developed exclusively on the HZTPH Dataset (N = 80) and evaluated on a completely independent external cohort from the JHSH Dataset (N = 230), which was not involved in training, validation, or parameter optimization. MSTANet achieved 93.61% accuracy on the development dataset and maintained 90.21% accuracy on the external validation cohort, exhibiting a minimal cross-center performance degradation. Overall, this study presents a robust and clinically feasible EEG-based depression detection framework, highlighting the translational potential of coarse-grained deep learning with an optimized dual-channel frontal EEG configuration for large-scale screening applications.