Age, Verbal IQ and Autism Severity Information Improves ASD Classification Based on Brain Morphometry

Background:

Previous findings of brain morphometry abnormalities in autism spectrum disorder (ASD) have been inconsistent (Chen 2011). In addition, previous large multi-site studies that use brain morphometry to classify ASD from typically developing controls (TDC) report low classification accuracies (<60 %) (Haar 2014; Katuwal 2015; Sabuncu & Konukoglu 2014). The variability of the reported findings and the low accuracies can be mainly attributed to the heterogeneity of ASD.

Objectives:  

In this study we investigate if the challenges posed by the heterogeneity of ASD brain morphometry for ASD classification can be improved by utilizing autism severity (AS), verbal IQ (VIQ) and age information.

Methods:  

Structural MRIs of 373 ASD and 361 TDC male subjects (age range: 6-40years) from the ABIDE were preprocessed using FreeSurfer. For each image, 538 brain morphometric features were derived. ASD vs. TDC classification was performed using the Random Forest classifier on morphometric features. Classification success was measured using the area under the curve (AUC) metric estimated by 10-fold cross validation. AS, VIQ and age information were utilized in two ways. First, VIQ and age information were used in conjunction with morphometric features. Second, subjects were divided into 3 sub-groups each for AS, VIQ and age (Figure 1). In this scheme, subject numbers of ASDs and TDCs were matched either by ‘up-sampling’ the smaller class in each training fold or by ‘down-sampling’ the larger class to match demographics of the smaller class.

Results:  

A moderate AUC of 0.61, similar to the best of previous work, was achieved when only brain morphometric features were used for classification. Adding age and VIQ to brain morphometry features improved the AUC to 0.68. When subjects were divided into sub-groups, classification performance improved significantly and AUC patterns of both the up-sampling and down-sampling schemes matched (Figure 1). The highest AUC of 0.92 was achieved in the down-sampling scheme for the low AS group (Figure 1B). The most important features for classification varied across sub-groups (Figure 2), however, they were predominantly from the left amygdala, right hippocampal, ventricular, insular, frontal and temporal regions. Further, the mean of Cohen’s d metric of the important features followed the AUC patterns. In both schemes, AUC was high for low AS (4 to 5) but decreased for moderate AS (6 to 7) and high AS (8 to 10). AUC decreased with VIQ. AUC was moderate for age groups of 6-13years and 18-40years but low for 13-18years age group. See Figure 1 for AUC values. 

Conclusions:  

The variability of the important features for classification across the sub-groups indicates that brain anatomical abnormalities in ASD are dependent on factors such as AS, VIQ and age. The increase in classification performance with the utilization of the above information demonstrates that the challenges posed by ASD heterogeneity can be mitigated by sub-grouping ASD. This study shows that the search for brain markers for sub-groups of ASD may be more fruitful than searching for markers across the whole spectrum of autism.