Depression Recognition Based on Pre-trained ResNet-18 Model and Brain Effective Connectivity Network

Depression has emerged as a primary health burden globally. Therefore, effectively identifying depression has become a significant challenge and obstacle in the field of public health. There are currently numerous issues with using electroencephalography (EEG) signals for depression identification. These issues include neglecting the dynamic characteristics of brain electrical signals, the information transmission relationships among them, and the low recognition accuracy of the features. Regarding the issues mentioned above, this study proposes an effective connectivity method based on transfer entropy to extract brain network features and achieved depression identification using pre-trained models. First, calculate the transfer entropy between all electrodes to obtain the corresponding matrix. Then, convert it into a three-dimensional RGB image, namely a heatmap. The created image is then used as input for pre-trained models, enabling the recognition of depression. In addition, to verify whether high-density EEG signals are beneficial for depression recognition, this study tested the recognition effect using 32-channel and 128-channel respectively. By comparing with other models, it was found that our model achieved the highest accuracy in depression recognition, with the highest accuracy rates achieved on 128 and 32 electrodes being 93.21% and 84.91%, respectively. Furthermore, we observed that the recognition rate of extracted EEG features was improved by 10% points in high-density space compared to low-density space. This indicates that high-density EEG is more effective for depression recognition. Overall, our study presents a novel approach and method for improving the accuracy of depression detection using EEG, which holds significant research implications.