Objectives: We sought to characterize alterations in local, regional, and global connectivity in ASD using resting-state fMRI, as well as develop a reliable method for classifying whether an individual had a diagnosis of ASD or was typically developing (TD) based on these measures.
Methods: Our sample consisted of 6-minute resting-state fMRI scans of eighty children and adolescents (43 ASD, 37 TD), matched for age (13.2+/-2.2), Full Scale IQ (104+/-12.3), gender (86% male), and head motion. Data were first corrected for motion, skull stripped, spatially smoothed, and temporally filtered. This was followed by regression of motion parameters, CSF, WM and whole-brain timeseries, and standard space registration. Residual gray matter timeseries were then used in voxelwise regional homogeneity (similarity of a voxel with those of its nearest neighbors using Kendall's W), global connectivity (average connectivity between a voxel and all other positively correlated voxels) and connectome-wide association (comparing a voxel's whole-brain connectivity map between participants) analyses. A multiple support vector machine recursive feature elimination algorithm (mSVM-RFE) was used to obtain a ranked list of voxelwise features for each of these methods. This process was wrapped in an external layer of 10-fold cross validation. Estimates of generalization error were obtained by averaging the performance of a tuned and trained radial basis function SVM classifier on the respective hold-out samples across these 10 folds. Top features from each method were then tested separately and jointly in an independent replication sample consisting of thirty subjects (15 ASD, 15 TD).
Results: Autism was associated with reductions in both local and global connectivity across multiple brain regions including the precuneus/posterior cingulate and medial prefrontal cortex of the default mode network, frontal and parietal regions of the attention network, as well as the striatum, anterior insula and fusiform gyrus. Within the training set, features derived from global connectivity, regional homogeneity, and the connectome-wide similarity matrix reached average accuracies of 70%, 74%, and 75% respectively. In the independent replication sample a classifier trained from these measures reached accuracies of 70%, 66%, and 66% respectively and 73% jointly (p=0.003; 11/15 ASD and 11/15 TD).
Conclusions: These findings support the notion that autism is characterized by reductions in both long and short-range functional connectivity across multiple brain networks. Furthermore, this work suggests that a 6-minute resting fMRI scan can distinguish individuals with autism above chance and may be useful as a diagnostic measure. Larger sample sizes, better data quality and analytical methods, as well as a better understanding of genetic factors influencing these circuits should improve accuracy.