TAU-DI Net: A Multi-Scale Convolutional Network Combining Prob-Sparse Attention for EEG-based Depression Identification

EEG-based detection of major depression disorder (MDD) plays a pivotal role in the subsequent treatment and recovery. With the rapid development of deep learning, CNN, LSTM, and attention-based models have been used for auxiliary diagnosis of MDD from EEG signals. However, these approaches either lack the utilization of pathological features of depression at the signal level or neglect handling redundant signal information within resting-state EEG. Considering these limitations, we propose a novel architecture based on an adaptive time-frequency distribution network. This method conducts the structure, which combines frequency-periodic transformation and multi-scale CNN to extract multi-frequency information representations from the most significant and least significant frequencies. Then, the method employs an adaptive weighted fusion of spatiotemporal representations across different frequencies and subsequently utilizes down-sampled Prob-Sparse Attention to distill reliable patterns within resting-state EEG. Experimental results demonstrate that our approach outperforms other models based on self-attention mechanisms and convolutional neural networks, achieving an optimal classification accuracy of 94.91%. The results show that adaptive methods based on different frequencies can produce better capabilities in processing EEG signals related to depression disorder.