Additional file 1 of Two neuroanatomical subtypes of males with autism spectrum disorder revealed using semi-supervised machine learning

Additional file 1. Table S1. Quality control metrics for the enrolled subjects. Table S2. Lists number of ASD subjects with available clinical measures. Table S3. List of brain regions in the Neuromorphometrics atlas used as features in HYDRA. Table S4. Detailed statistical analysis of age between HC and ASD in each site and full cohort. Table S5. Clinical profiles between subgroups across the 7 sites. Fig. S1. The image quality measures describe the properties of the image. Fig. S2. Examples of different quality ratings for the original T1 images (left) and segmented images (right). Fig. S3. Schematic illustration of the HYDRA method. Controls (denoted by blue squares) are separated from the patients (denoted by red triangles) using a convex polytope decision boundary. Solid lines correspond to the classifier, dashed lines indicate the margin while highlighted linear segments define the separating convex polytope. Fig. S4. Cross-validated stability of ASD subtypes: Adjusted Rand Index (ARI) vs. number of subtypes (K) indicating highest reproducibility when K = 2. Fig. S5. Volume difference in gray matter (A) and white matter (B) volume between healthy control (HC) (n = 257) and ASD (n = 221) by standard case-control comparison. Effect size (Cohen���s d) maps were generated from regional volumetric maps masked by the set of regions that showed statistically significant differences (PFDR < 0.05) in the MIDAS analysis. Fig. S6. GM volumetric differences between each subtype and HC for K = 2 in Split 1 (left column) and Split 2 (right column). Fig. S7. WM volumetric differences between each subtype and HC for K = 2 in Split 1 (left column) and Split 2 (right column). Fig. S8. The number of overlaps assigned to the same ASD subtype in the leave-one-site-out strategy. Fig. S9. Temporal SD of dynamic R-fMRI measures differences between ASD 1 and ASD 2 using the two-sample t-tests for (a) DC, (b) GSCorr and (c) ReHo (GRF, voxel-level p < 0.001, cluster-level p < 0.01, two-tailed). Fig. S10. (a) Two sample t-test of the mean of concordance among R-fMRI indices between ASD 1 and ASD 2; (b) Two sample t-test of the SD of concordance among R-fMRI indices between ASD 1 and ASD 2. Of note, the demonstrated mean/SD were fitted values with the effect of head motion, age and site regressed out. Fig. S11. Temporal SD of dynamic stability measures differences between ASD 1 and ASD 2 using two-sample t-test (GRF correction at voxel-level p < 0.001, cluster-level p < 0.01, two-tailed). Fig. S12. Temporal SD of dynamic R-fMRI measures differences between healthy controls and ASD 1 using the two-sample t-tests for (a) DC, (b) GSCorr and (c) ReHo (GRF, voxel-level p < 0.001, cluster-level p < 0.01, two-tailed). Fig. S13. (a) Temporal SD of dynamic stability measures differences between healthy controls and ASD 1 using two-sample t-test (GRF correction at voxel-level p < 0.001, cluster-level p < 0.01, two-tailed). (b) Temporal SD of dynamic stability measures differences between healthy controls and ASD 2 using two-sample t-test (GRF correction at voxel-level p < 0.001, cluster-level p < 0.01, two-tailed). Fig. S14. Average weight ranking differences of ROI in the subtype-healthy control classification. Subtype1-healthy controls (Left) and subtype2-healthy controls (Right). The average weight of the top 40 ROIs is shown here.