Disrupted functional connectivity patterns have been increasingly used as features in pattern recognition algorithms to discriminate neuropsychiatric patients from healthy subjects. Deep neural networks (DNNs) were employed to fMRI functional network classification only very recently and its application to EEG-based connectome is largely unexplored. We propose a DNN with deep belief network (DBN) architecture for automated classification of schizophrenia (SZ) based on EEG effective connectivity. We used vector-autoregression-based directed connectivity (DC), graph-theoretical complex network (CN) measures and combination of both as input features. On a large resting-state EEG dataset, we found a significant decrease in synchronization of neural oscillations measured by partial directed coherence, and a reduced network integration in terms of weighted degrees and transitivity in SZ compared to healthy controls. The proposed DNN-DBN significantly outperforms three other traditional classifiers, due to its inherent capability as feature extractor to learn hierarchical representations from the aberrant brain network structure. Combined DC-CN features gives further improvement over the raw DC and CN features alone, achieving remarkable classification accuracy of 95% for the theta and beta bands.
|Original language||English (US)|
|Title of host publication||2019 9th International IEEE/EMBS Conference on Neural Engineering (NER)|
|Publisher||Institute of Electrical and Electronics Engineers (IEEE)|
|Number of pages||6|
|State||Published - Mar 2019|