Deep learning applied to 4-electrode EEG resting-state data detects depression in an untrained external population

In this study we trained and tested several deep learning algorithms to classify depressive individuals and controls based on their electroencephalography data. Traditionally, classification methods based on electroencephalography resting-state are based primarily on linear features or a combination of linear and non-linear features. Based on different theoretical grounds, some authors claim that the more electrodes, the more accurate the classifiers, while others consider that working on a selection of electrodes is a better approach{square}. In this study, a data-driven approach was initially applied on a selection of electrodes to classify 25 depressive and 24 control participants. Using a classifier with just four electrodes, based on non-linear features with high temporo-spatial complexity, proved accurate enough to classify depressive and control participants. After the classifier was internally trained and tested, it was applied to electroencephalography resting-state data of control and depressive individuals available from a public database, obtaining a classifier accuracy of 93% in the depressive and 100% in the control group. This validates the generalizability of the classifier to untrained data from different teams, populations and settings. We conclude that time-window span analysis is a promising approach to understand the neural dynamics of depression and to develop an independent biomarker.