Examining Associations between Brain Morphology and Social Function in ASD, ADHD, OCD, and Typical Development Using Machine Learning: Analysis of Pond Network Data

Oral Presentation
Saturday, May 12, 2018: 2:52 PM
Jurriaanse Zaal (de Doelen ICC Rotterdam)
A. Kushki1,2, M. Komeili3, S. Panahandeh4, E. Anagnostou5 and J. P. Lerch6, (1)Bloorview Research Institute, Toronto, ON, Canada, (2)Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, ON, Canada, (3)University of Toronto, Toronto, ON, Canada, (4)Institute of Biomaterials and Biomedical Engineering, University of Toronto, Vaughan, ON, Canada, (5)Holland Bloorview Kids Rehabilitation Hospital, Toronto, ON, Canada, (6)Mouse Imaging Centre, Hospital for Sick Children, Toronto, ON, Canada
Background: There is significant variability in biology and behavior within the autism spectrum. At the same time, autism spectrum disorder (ASD) shares biological and behavioral traits with other neurodevelopmental disorders including attention deficit/hyperactivity disorder (ADHD) and obsessive compulsive disorder (OCD). This challenges existing diagnostic categories and motivates a diagnosis-agnostic and data-driven approach to characterizing variability across the autism spectrum and related neurodevelopmental disorders.

Objectives: The objectives of this study were two-fold: 1) characterize variability in brain-behaviour associations across ASD, ADHD, OCD, and typical development (TD) using a data-driven and diagnosis-agnostic approach, and 2) examine whether or not diagnostic labels are associated with distinct brain-behaviour patterns.

Methods: Data from a sample of 218 participants in the POND Network studies were used for analysis (nASD=104, age:11.3±2.7; nADHD=62, age:10.9±2.5; nOCD=35, age:12.1±2.3; nTD=17, age:10.9±2.7). Brain data included cortical thickness measurements from 76 regions of the brain obtained using the CIVET pipeline, and corrected for total brain volume, age, and site. Behavioural data were total scores on the Social Communication Questionnaire (SCQ). Analyses were performed using a novel machine learning pipeline comprising two steps: 1) a feature selection step, using sequential feature forward selection, to determine the brain regions whose variability best aligned with that of the behavioural data, and 2) discovery of clusters that were aligned across brain and behaviour data using co-regularized multi-view spectral clustering. To ensure stability of the found patterns, the analyses were run on 500 random partitions of the data, each including 90% of the participants. Only participants who were assigned to the same cluster over 50% of the time (significantly higher than chance) were considered connected in the final analysis.

Results: The results revealed a complex pattern of brain-behaviour association, characterized by a many-to-many mapping. Five groups emerged based on clustering. Cluster boundaries were not crisp for groups 2-5. Groups generally contained participants from all diagnostic categories (% dx (ASD:ADHD:OCD:TD): group 1 - 37:3:6:0, group 2- 23:19:9:6, group 3- 35:18:6:6, group 4- 4:29:34:29, group 5- 2:31:46:59). The groups differed significantly on SCQ scores and IQ (p<0.0001), but not age. Group 5 had the lowest SCQ score and the highest IQ, whereas group 1 had the highest SCQ and lowest IQ. There were also significant group differences in cortical thickness in the left lingual gyrus and the right precuneus, the two regions that best aligned the brain and behavioural data (p<0.0001). Three groups with the largest proportion of ASD participants (35%, 37%, and 23%) showed varying SCQ scores (16.5+4.4; 26.9+3.1; 13.9+4.7) and cortical thickness values (Lingual gyrus (z-score): 0.4+0.1; -0.2+1.0;-0.7+0.7).

Conclusions: Examining brain and behavioural data from a sample of children with neurodevelopmental disorders, this study revealed a complex and many-to-many association among cortical thickness and social function measured by the Social Communication Questionnaire. The results suggests that brain-behaviour patterns are shared among ASD and ADHD, OCD, and typical development, supporting biological and behavioural overlap among these disorders, as well as a a dimensional model of traits that extends into typical development.