Multi-Class Pattern Classification Discriminates Young Males and Females with Autism Spectrum Disorder Based on Cortical Morphology

Background: Autism spectrum disorder (ASD) is diagnosed much more frequently in males than in females. One prevailing theory for explaining the male bias in ASD is the multifactorial liability model, which suggests that multiple liability factors shift the vulnerability threshold for ASD higher in females, such that a smaller proportion of females reach clinical diagnosis. One prediction from this model is that females with ASD will exhibit a higher liability of neural alterations and will deviate further from typically developing (TD) females than their male counterparts. There is growing evidence supporting distinct pattern of neural alterations in females with ASD in older individuals, but studies of very young children are lacking.

Objectives: We utilized cortical surface area, thickness, and volumetric measurements to evaluate within- and between-sex and diagnosis differences in a large cohort of preschool-aged children with ASD and TD controls.

Methods: We acquired structural T1-weighted MRIs in 213 children with ASD (150 male, 63 female) and 101 age-matched TD controls (54 male, 47 female). Mean age at time of MRI acquisition was 39 months. Cortical reconstruction and volumetric segmentation were performed within FreeSurfer v5.1.0. Cortical gray matter volumes, cortical thicknesses and surface areas were extracted for 68 gyral regions (Desikan-Killiany atlas). A multi-class linear discriminant model was used to identify patterns of brain regions that discriminate between sex and diagnosis (ASDf, ASDm, TDf, TDm). The validity of the model was tested using a leave-four-out cross-validation (leave one sample from each and every of the four groups out) to avoid overfitting with biased sample size. This process also allowed the training and test cases to remain independent. Finally, permutation analysis was performed to empirically determine if the obtained classification accuracy was significantly greater than chance. We report: (1) classification accuracy of test sets, (2) feature weights that contributed to the classification, and (3) the overlap and non-overlap in brain regions across different comparisons.

Results: Our preliminary results suggest that the multi-class linear discriminant model could achieve an accuracy of 47.2% which was significantly higher than chance (p = 0.001, null accuracy = 27.4%). Interestingly, the two most-discriminant vectors for the multiclass model distinguished sex and diagnosis, respectively (Figure 1). Regions specific to diagnosis comparison (ASD vs. TD) included regions in neural systems related to understanding mental states of others, expressive language and reception of facial communication (right temporal pole, bilateral lateral occipital, bilateral fusiform) as well as bilateral transverse temporal, left entorhinal, and superior parietal lobule. Regions specific to sex comparison included right pars triangularis, posterior cingulate, superior temporal sulcus, and left lingual gyrus. Regions overlapping across sex and diagnosis comparison included bilateral middle frontal gyurs, left rostral anterior cingulate, precuneus, and inferior temporal gyri. Surface area measures contributed the most to the classification.

Conclusions: These preliminary data suggest that females and males with ASD have patterns of cortical alterations that are distinguishable from TD males and females. In addition, neural patterns of diagnostic differences are more pronounced in females than in males, providing support for the multifactorial liability model.