Using deep learning to decode abnormal brain neural activity in MDD from single-trial EEG signals

Abstract

Objectives The application of electroencephalography (EEG) to the study of major depressive disorder (MDD) is a common approach. However, there is no one-to-one correspondence between EEG and brain neural activity, and it is unclear whether single-trial EEG signals detect the cognitive neural activity of MDD. Methods Here, we used deep learning to explore this issue. Deep learning adopted in this paper was an end-to-end classification method named EEGNet which could update model parameters automatically based on the characteristics of the data and classify MDD and healthy subjects (HS) from single-trial EEG signals to obtain classification results above the chance level. Furthermore, the saliency map was used to analyze the neural network model and visualize the channels and time periods that contributed the most. Results Finally, EEGNet achieved an average classification accuracy of 61.4% in the four categories, and the result of feature visualization was consistent with the cognitive neural interpretation of existing studies. Conclusion The findings suggested that deep learning could help cognitive neuroscience explore neural activity.

Keywords

• eeg
• deep learning
• mdd
• neural activity
• saliency map