Characterizing replicability in the clustering structure of brain morphology in autism,...

In neurodevelopmental research, within-diagnosis heterogeneity and across-diagnosis overlap necessitate a shift from case-control designs to data-driven clustering approaches. However, our understanding of the replicability of these clustering structures across independent datasets remains limited. Our objective was to examine the replicability of clustering structure in measures of brain morphology in neurodiverse children across two independent datasets, namely the Province of Ontario Neurodevelopmental Disorder (POND) Network and the Healthy Brain Network (HBN). POND and HBN data were collected across various institutions in Ontario, Canada, and New York, United States, respectively. Participants were 5-19 years old and had diagnoses of autism, attention deficit/hyperactivity disorder (ADHD), obsessive compulsive disorder (OCD), or were neurotypical. We used measures of cortical volume, surface area, cortical thickness, and subgroup volume from structural MRI data. Principal component analysis (PCA) and clustering were used to examine the replicability of clustering structures across the datasets. Correlations among principle components, measures of clusterability, and alignment between the four brain measures as well as male/female subsets were examined. Brain-behaviour associations were examined using univariate and multivariate approaches. The POND dataset included 747 participants with (autism n = 312, ADHD n = 220, OCD n = 70, neurotypical n = 145). The HBN dataset included 582 participants (autism n = 60, ADHD n = 445, OCD n = 19, neurotypical n = 58). Our results showed significant between-dataset correlations in 82.1% of the principal components derived from brain measures. A two-cluster structure was replicated across datasets, brain measures, and the female/male subsets, however the participant composition of clusters were only aligned between cortical volume and surface area, and cortical thickness and subcortical volume. Regional effect sizes for between-cluster differences were highly correlated across datasets (beta = 0.92+/-0.01, p < 0.0001; adjusted R-squared=0.93). Data-driven clusters did not align with diagnostic labels across datasets. Brain-behaviour associations were only replicated for male subsets and subcortical volume using multivariate analysis. We found evidence of replicability of the clustering structure across two independent datasets; however, caution must be exercised in integrating multiple measures in clustering and interpretation of brain-behaviour associations.