adonunes/ct_trajectory_features_asd_2019

Age-related changes in Cortical Thickness in Autism

Cort_Thick_age-changes_Autism_2018

Age-related changes in Cortical Thickness in Autism

This set of scprips takes cortical thickness (CT) values from the ABIDE repository and :

[ in S01_getGroups.m ]

• Excludes outliers per center.
• Creates a group of ASD and TD, balanced by number of subjects, age and the ratio male/female per center.

[ in S02_ModelFitting.m ]

• Center variability is taken out while accounting for group, age and age-group interactions in a linear, quadratic and cubic models. In the following steps, CT values corrected for center variability are used.
• Significance of a model fit using a deviance test is measured for each area of the CT parcellation, areas that do not fit on a group level are masked out. ASD and TD are tested separately.

[ in S03_subsampling_analysis_Hemi.m & S03_subsampling_analysis_Atlases.m ]

• For each group, subsamples of subjects are created. This subsamples have the flattest age distribution possible.
• For each subgroup, a linear, quadratic and cubic models are fitted and the highest order coefficient (linear, quadratic and cubic trends) are extracted for statistical testing between groups. These coefficients are invariant across age.
• Using Partial Least Squares (PLS), the coefficients at each area of teh parcellation are tested for significance.

[ in S04_subsampling_ADOS_corr.m ]

• The coefficients derived from ASD subgroups and the mean ADOS scores from the subgrup of subjects are tested for correlations. Each model is correlated with mean ADOS with PLS, the p-vals are bonferroni corrected.
• The three models were analyzed separately, for simplicity, as only one highly correlated.

[ in S05_subsampling_SVM_classification.m ]

• A support vector machine(SVM) was trained and tested with 500 cross-validations.
• For each run, the ASD and TD groups where splitted in half. In the training set, coefficients were extracted using one parcellation, and for each model coefficient a SVM was trained to classify age-related changes velonging to ASD or TD samples. Then, with the other dataset, the accuracy, sensistivity and specificity was tested.