A Novel Severity Measure for Quantitative Description of Heterogeneity in Autism

Background:  The evaluation of individuals with ASD is usually assessed by phenotypic scores describing different aspects of the clinical presentation.  ASD heterogeneity, however, makes detection of differences between affected individuals and typically developing controls (TDCs) difficult when phenotypic measures are imprecise and lead to overlap for ASD and TDC samples. Many research applications would benefit greatly from a continuous severity measure to characterize the impact of specific etiologically significant factors, including genetic risk variants and brain processes captured by neuroimaging data.

Objectives:  The aim of this work is to design a severity measure that quantifies individual differences in the autism spectrum by fusing information from various phenotypic scores that describe different aspects of functioning, when individually none provide a complete characterization. The validity of the measure will be tested for its ability to categorize sample differences in diffusion tensor brain imaging data.

Methods:  A dataset of 370 male subjects with ASD and 118 male TDCs was collected. From a large battery of measures, 50 key phenotypic scores were chosen for the analyses. The individual severity belief of a score was estimated from the cumulative distribution function of the score defined over the ASD sample. The weighting of these individual decisions was established based on the average classification error between ASD and TDC when using these scores. Probabilistic voting was used to fuse individual decisions of different phenotypic scores in a weighted manner. The measure was used to determine groups of high and low severity in the population. These subgroups were then used to study imaging-based group differences on DTI data acquired with 30 gradient directions. Fractional anisotropy (FA) maps computed from the DTI data.

Results:  When the ASD sample is stratified according to this severity measure, voxel-wise analysis of the FA maps revealed differences that could not be seen when the sample was analyzed as a whole. Regions of differences agree with previous diffusion-based studies in ASD. Several regions, including the body of corpus callosum, corona radiate, superior longitudinal fasciculus and the inferior fronto-occipital fasciculus show significant group differences only when TDCs are compared to the more severe tail of the ASD distribution that is defined by this severity measure. These results are compared to stratification by a traditional dimensional measure of ASD – the Social Responsiveness Scale (SRS), which focuses on the core diagnostic symptoms. By fusing a broad range of scores, many of which are ancillary to the core of ASD, this new severity metric performs as well or better.

Conclusions:  We have designed a severity measure that combines information from individual phenotypic scores to define a measure of heterogeneity. It does better in detecting TDC vs ASD differences than any single test scores. DTI-based imaging analysis of groups selected based on this severity score revealed larger differences between ASD and TDC samples.