Classification of autism spectrum disorder from resting-state fMRI with mutual connectivity analysis

Abstract

In this study, we investigate if differences in interaction between different brain regions for subjects with autism spectrum disorder (ASD) and healthy controls can be captured using resting-state fMRI. To this end, we investigate the use of mutual connectivity analysis with Local Models (MCA-LM), which estimates nonlinear measures of interaction between pairs of time-series in terms of cross-predictability. These pairwise measures provide a high-dimensional representation of connectivity profiles for subjects and are used as features for classification. Subsequently, we perform feature selection, reducing the dimension of the input space with the Kendall’s τ coefficient method. The Random Forests (RF) and AdaBoost classifiers are used. Performing machine learning on functional connectivity measures is commonly known as multi-voxel pattern analysis (MVPA). Traditionally, measures of functional connectivity are obtained with cross-correlation. Hence, as a metric to evaluate MCA-LM against, we also investigate classification performance with cross-correlation. The high area under receiver operating curve (AUC) and accuracy values for 100 different train/test separations across both classifiers using MCA-LM (mean AUC ranges between 0.78 - 0.85 and mean accuracy between 0.7 - 0.81) compared with standard MVPA analysis using cross-correlation between fMRI time-series (mean AUC ranges between 0.54 - 0.6 and mean accuracy between 0.50 - 0.57), across all the number of features selected demonstrates that such a nonlinear measure may be better suited at extracting information from the time-series data and has potential for the development of novel neuro-imaging biomarkers for ASD.