Fig. 1.

Outline of methods for brain-age prediction using machine learning.

The data includes a training set and 2 testing sets (cohort1: controls vs. PWS; cohort 2: controls vs. SNORD116 microdeletion). The data was pre-processed using the Statistical Parametric Mapping software (SPM). The T1 images underwent tissue segmentation after initial quality control, separating GM, WM and CSF. The segmented images were then normalized using DARTEL for nonlinear registration, then resampled to the MNI152 template using a 4 mm smoothing kernel. GM and WM maps were concatenated into a single vector of data relating to brain size. All data underwent voxel-wise similarity analysis to generate a similarity kernel using PRoNTo. The training set was used to generate the brain-age model via supervised machine learning that produced a Gaussian Processes Regression (GPR) model trained to recognize structural patterns from imaging data associated with chronological age. The accuracy of the model generated from the training set was assessed using a 10-fold cross-validation method whereby 10% of samples were used for testing in all possible iterations to generate age predictions on all samples. The trained and validated GPR model was applied to the two test groups.