SVM Classifies Age but Not Autism Risk Using fcMRI Data from 6 and 12-Month-Old Infants at Low and High Genetic Risk for Autism

Background:  Human large-scale functional brain networks are hypothesized to undergo significant changes during development. Little is known about the timing of these functional architectural reorganizations, and there have only been a few infant functional brain imaging studies. Other recent and seminal reports of developmental change in functional brain networks have been called into question because of appreciation of motion artifact effects on resting-state functional connectivity magnetic resonance imaging (fcMRI) data.

Objectives:  To address an important part of this developmental question, we made multivariate pattern classifications of fcMRI data acquired in on-going, multi-site, longitudinal studies of brain and behavioral development in infants at low and high genetic risk for autism spectrum disorder (ASD).

Methods:  The sample included six and 12-month low- (control) and high-risk (older sibling with ASD) infants not meeting clinical best estimate criteria for ASD at 24 months. fcMRI data were processed according to recent analytic and motion cleaning recommendations (Power et al., 2013) with infant-specific adaptations to initial registration and nuisance regression steps. fcMRI matrices were constructed using 230 (from 264 in Power et al., 2011 plus 16 additionally derived from Philip et al., 2012) functionally-defined seed regions which trained image analysts agreed were appropriately positioned in grey matter at both ages. Support vector machine (SVM) methods involved recent adaptations of those used by some of the authors (Dosenbach et al., 2010). 64 datasets from each risk group were pseudo-randomly selected from a total of 164 to allow for balanced, two-group classifications. SVM steps included t test filtering to 200 features, linear kernel, soft margin separation, and leave-one-out-cross-validation (within group). Data from low-risk subjects were then classified with the high-risk-trained SVMs and vice-versa. Control tests accounted for mixed cross-sectional and longitudinal data groupings. Results from data pre-processed using age-specific atlases were compared against results from the same data pre-processed using a cross-age target atlas intermediary. Sensitivities and specificities were measured, and significance was assessed in reference to binomial probabilities. The classification vector from a 128 dataset run allowed visualization of contributing functional connections and seeds.

Results:  SVMs classified six versus 12 month-old infants based on fcMRI data, alone: high-risk – accuracy = 75%, sensitivity = 81.3%, specificity = 68.8%, p = 1.22e-05; low-risk – accuracy = 81.3%, sensitivity = 78.1%, specificity = 84.4%, p = 5.03e-08. SVMs could not classify genetic risk for ASD at either age – all p > 0.05. The classification vector illustrates contributions from functional connections between regions which in adultswould populate default mode, somatosensory-motor, cingulo-opercular, and visual networks. Though not exclusive, weights show a pattern whereby stronger functional connections for longer-distance, anterior-posterior functional connections contribute more to classification of 12 months, and those for posterior, left-right functional connections contribute more to classification of six months.

Conclusions: Results support the developmental hypothesis of significant change in functional brain organization with age, here, during a six month period in the first year of life. Findings encourage other basic developmental examinations and pursuit of questions about functional brain network differences associated with ASD.