31957
Using Canonical Correlation Analysis Approaches to Develop a Circuit-Based Clinical Measure of Autism Core and Associated Symptoms
Key challenges for clinical trials of autism include the clinical and etiological heterogeneity of the condition, and the lack of sensitive outcome measures. Existing clinical measures typically assess a collection of autism-related clinical symptoms. However, it is likely that various behavioural symptoms – even within specific domains – may have different neurocognitive or neurobiological underpinnings. Moreover, due to the diversity of autism the same behavioural symptom may result from different neurobiological abnormalities in different autistic individuals. Therefore, purely behaviourally-defined measures that encompass various neurobiological processes may not be sufficiently sensitive to detect treatment effects. Here we adopt a novel approach starting with robustly defined brain circuits that underpin fundamental social, emotional, motivational, and cognitive processes, and then mapping brain-behaviour relationships to different subgroups.
Objectives: To create a “circuit-based clinical outcome measure” we used Canonical Correlation Analysis (CCA) based approaches to identify items from existing clinical measures that best covary with brain function in robust, pre-defined brain circuits, and explored whether these brain-behaviour relationships vary between different autism subgroups
Methods:
Combined clinical-fMRI data sets were available from ~100-300 autistic individuals and 80-220 control participants from the EU-AIMS Longitudinal European Autism Project. We used two CCA-based methods that allow us to identify relationships between two multivariate datasets (here clinical items and fMRI regions-of-interests) using information from cross-covariance matrices and that employ internal cross-validation mechanisms to enhance the robustness of the fitting. Regularised CCA can handle more samples than participants when required and maximise the correlation between clinical items and brain regions. To estimate the robustness of the canonical scores we used bootstrapping to generate clinical and imaging variable loadings on a subset of participants and then used these variable loadings to calculate median values of the estimated canonical scores on the excluded subset.
Canonical Vector Regression (CVR) uses variable loadings from CCA to obtain a regression fit to group membership (TD/ASD). We then entered individual scores from both methods into a cluster analysis using fuzzy c means and affinity propagation-based clustering with bootstrap estimation of the optimal cluster number. By combining CVR and RCCA we simultaneously separated individuals on two complementary aspects of the data - maximal correlation between brain and clinical symptoms and maximal distance between groups.
Results:
Preliminary analyses revealed robust correlations between network function and clinical symptoms between autism ‘subgroups’. For example, RCCA and CVR between the Strength and Difficulties Questionnaire (SDQ) and ROIs in a reward task showed significant correlations in the first canonical variate (RCCA p=.001, CVR p=.024). 4 items significantly covaried with changes in ROIs; while ventral striatal activity in two conditions significantly covaried with changes in SDQ.
We found four observed clusters; two of which largely separated an ASD-subgroup from controls (Fig 1a) and maintained high predictive accuracy based on the predicted output (Fig 2).
Conclusions:
These CCA-approaches can identify clinical items that differentially covary with network function in ASD “subgroups”. Next, we will sort clinical items according to network function across four fundamental bio-behavioural domains, and ‘harmonise’ them in terms of response format.