Replicable Network-Based Diagnostic Classification of ASD in the Autism Brain Imaging Data Exchange

Background:  Autism spectrum disorders (ASDs) are characterized by alterations in and heterogeneous patterns of functional brain connectivity. This heterogeneity has presented a significant obstacle to MRI/fMRI-based diagnostic classification because most optimization algorithms assume a single, mutually exclusive distinction between control and ASD groups. As a solution, in this study we present a novel classification method based on a combination of network-based effective connectivity (structural vector autoregression; SVAR) and a high-dimensional variable selection procedure (smoothly clipped absolute deviation; SCAD) to identify subsets of connections that are robust to high-dimensional heterogeneity, and as such may have promise when applied to fMRI data in order to predict ASD diagnostic status.

Objectives:  Determine the efficacy of a high-dimensional variable selection procedure (smoothly clipped absolute deviation; SCAD) in determining diagnostic status (ASD versus non-ASD) in a large, publicly available repository of resting-state data (Autism Brain Imaging Data Exchange; ABIDE), and replicate the classification procedure across 16 physical ABIDE sites. 

Methods:   We used resting state-fMRI data from 1,112 subjects (ASD N=539, Control N=573) across 16 geographic sites from ABIDE, to measure functional connectivity estimates among 12 regions of interest (ROIs) in the frontostriatal circuit. We incorporated all 12 ROIs into subject-level network maps, yielding 144 directed connectivity paths per subject. Using a leave-one-out procedure, we compared predicted diagnosis from SVAR+SCAD against actual diagnosis based on (ADOS/ADI-R). 

Results:  Results indicated that fronto-striatal network maps contained, at the individual- and site-levels, sufficient data to accurately classify 84.6% of cases, on average, across all 16 ABIDE sites. Correlational and temporally-lagged models performed relatively poorly, whereas a model that combines instantaneous and temporally-lagged effects (structural vector autoregression; SVAR) provided a superior model fit, achieving greater than 90% accuracy for sites with less than 50 subjects.

Conclusions:  Network-level information that is variably expressed as the directional connective properties of the fronto-striatal circuit can be captured by SVAR+SCAD and used to predict the diagnosis of ASD via a leave-one-out prediction algorithm. Although no significant associations between any single connectivity estimate and ASD diagnosis were observed, SVAR+SCAD captured network-level properties that are not currently incorporated into other machine-learning and prediction algorithms, which focus on a low- (or one-) dimensional distinction between groups. Replication of predictive estimates across ABIDE sites increases confidence in the utility of this method for fMRI-based classification enterprises.