Classification of brain functional connectivity using convolutional neural networks

Abnormalities and alterations in brain connectivity networks as measured using neuroimaging data has been increasingly used as biomarkers for various neuropsychiatric disorders. Schizophrenia (SCZ) is a complex neuropsychiatric disorder associated with dysconnectivity in brain networks. In this paper, we develop a framework for automatic classification of healthy control and SCZ patient based on electroencephalogram (EEG) connectivity and compare the classification performance with conventional artificial neural network (ANN). We propose to use convolutional neural network (CNN) for the classification of brain functional connectivity between healthy control and SCZ groups. Vector autoregression (VAR) model is used to extract connectivity features from schizophrenia EEG signals and directed connectivity at different EEG frequency bands is computed via partial directed coherence (PDC). Results show that the classification with high accuracy is achievable using VAR model. From the result, the performance of CNN reaches 86.9% over five-fold cross validation that considered to be good accuracy for the CNN to do a good prediction. The results also show that time-domain VAR features performed better than frequency domain PDC features. CNN provides a more practical method in classification between healthy and schizophrenic brain connectivity.