RESUMEN
Objective: To explore the effectiveness of using deep learning network combined Vision Transformer (ViT) and Transformer to identify patients with depressive disorder on the basis of their sleep electroencephalogram (EEG) signals. Methods: The sleep EEG signals of 28 patients with depressive disorder and 37 normal controls were preprocessed. Then, the signals were converted into image format and the feature information on frequency domain and spatial domain was retained. After that, the images were transmitted to the ViT-Transformer coding network for deep learning of the EEG signal characteristics of the rapid eye movement (REM) sleep and non-rapid eye movement (NREM) sleep in patients with depressive disorder and those in normal controls, respectively, and to identify patients with depressive disorder. Results: Based on the ViT-Transformer network, after examining different EEG frequencies, we found that the combination of delta, theta, and beta waves produced better results in identifying depressive disorder. Among the different EEG frequencies, EEG signal features of delta-theta-beta combination waves in REM sleep achieved 92.8% accuracy and 93.8% precision for identifying depression, with the recall rate of patients with depression being 84.7%, and the F0.5 value being 0.917±0.074. When using the delta-theta-beta combination EEG signal features in NREM sleep to identify depressive disorder, the accuracy was 91.7%, the precision was 90.8%, the recall rate was 85.2%, and the F0.5 value was 0.914±0.062. In addition, through visualization of the sleep EEG of different sleep stages for the whole night, it was found that classification errors usually occurred during transition to a different sleep stage. Conclusion: Using the deep learning ViT-Transformer network, we found that the EEG signal features in REM sleep based on delta-theta-beta combination waves showed better effect in identifying depressive disorder.