Classification of autism spectrum disorders individuals and controls using phase and envelope features from resting-state fMRI data

This study uses phase and envelope features extracted from fMRI data for discriminating between Autism Spectrum Disorders (ASD) and control subjects. We exploit the transfer function perturbation (TFP) method to estimate the instantaneous phase and envelope of intrinsic resting-state brain network components from fMRI data. Then we calculated power, entropy, and coherency features. We examined three different classifiers and two different feature selection algorithms, in a way that the subsets of features that best predict classes were selected using a sequential forward feature selection (SFFS) algorithm and principal component analysis (PCA). Afterwards, three different categories of calculated features, including phase features, envelope features, and a combination of phase and envelope features, fed into non-linear Support Vector Machine (SVM), K-Nearest Neighbours (K-NN), and Deep Neural Network (DNN). Results illustrate that phase features are significantly discriminative and considerably improve ADSs and control subjects' classification accuracy and lead to robust prediction. Moreover, 91% of classification accuracy was obtained when the dimension of phase features was reduced by PCA and fed to the non-linear SVM. Eventually, the two-sample t-test and Pearson's correlation coefficient illustrated that phase features had a significantly lower correlation than envelope features.