“Deep learning, psychiatry and EEG: What we know and where to go”
Within recent years, artificial intelligence, and particularly deep learning (DL) techniques have revolutionized medical labelling tasks for diagnostic and predictive purposes in many fields. Since multichannel neurophysiological data, i.e. electroencephalogram (EEG) data can be arranged as two-dimensional matrices, they are well suited for DL analysis. The combination of high sampling-rate EEG data with advanced DL analysis techniques has the power to let the field of neurophysiology thrive.
This talk will give an overview on the current state of DL in neuropsychiatric disorders using EEG as input and clinically relevant parameters such as response or diagnosis as output. Hard-and software frameworks will be introduced, and possible caveats will be highlighted. Furthermore, DL for classification of some ground truth scenarios in large EEG datasets will be presented, including results for classification of subjects with high risk for psychosis. To line out future perspectives of DL analysis, the extraction of meaningful features from the DL networks trained on EEG will be explained, tracing back the DL-learned featured to spatial and functional patterns. This will allow us to gain more insight into pathophysiological mechanisms of neuropsychiatric disorders.