Using functional magnetic resonance imaging (fMRI) data, we have previously identified specific brain areas in which children with autism spectrum disorder (ASD) exhibit hypoactivation in response to biological motion compared to their typically developing (TD) counterparts (Kaiser et al., 2010, PNAS). Follow-up studies using limited sample sizes have shown that such state regions have promising utility as a diagnostic marker (Kaiser & Pelphrey, 2011, DCN).
Objectives:
In order to build upon our preliminary study, we aim to evaluate the power of the biological motion paradigm as a diagnostic tool by applying it to a full Replication cohort of children with and without ASD.
Methods:
Our analysis included a Discovery cohort with 17 ASD children (mean age = 10.93±3.09 years) and 22 TD children (mean age = 12.51±3.67 years) and a large Replication cohort with 37 ASD children (mean age=11.26±3.34) and 38 TD children (mean age=11.52 ±2.91). Children with ASD were diagnosed using ADOS, ADI-R, and expert clinical judgment. Subjects watched coherent and scrambled point-light displays of biological motion during fMRI acquisition. Individual general linear model T-maps (biological > scrambled motion) were computed. Using a whole-brain clustered subsampling approach, we identified regions-of-interests in which a support vector machine classifier that was modeled using Discovery cohort T-maps could predict the diagnosis of the children in the Replication cohort. Finally, we tested the effect of Replication cohort size on the reported accuracy of the classifier model by varying the number of children included.
Results:
The classifier obtained a peak sensitivity (percentage of children correctly identified as ASD) of 78%, and a specificity (percentage of correctly classified typically developing children) of 71% for a total classification accuracy of 75%. The most predictive clusters were located in the right fusiform gyrus, corroborating a previously defined state region (Kaiser et al., 2010, PNAS). The classification accuracy varied dramatically as a function of Replication cohort size, with classification accuracies over 85% for 30 subjects and less; for 60 or more subjects, however, the classification accuracy stabilized around 75%.
Conclusions:
We replicated our finding that biological motion processing is abnormally processed in the right fusiform gyrus in children with ASD. In addition, we demonstrated that it is feasible to use Discovery cohort data in this region to construct an ASD classifier that can distinguish between subjects with and without ASD in a large Replication cohort. Our study demonstrates the robustness of fMRI as a predictor of ASD and paves the path for future use of fMRI as an early diagnostic tool.