Fig. 1.

Sketch of the data analysis procedure. the MRI data of patients with autism spectrum disorder (ASD) and matched control subjects are preprocessed with the Statistic Parametric Mapping (SPM) software package to obtain the segmented grey matter (a), white matter (b), cerebrospinal fluid (c); the statistical analysis is performed on global volumes; the GM tissue of each subject is rolled down into a vector of features (d) and processed by a machine-learning analysis based on support vector machines (SVMs); to reduce the data dimensionality the recursive feature elimination (RFE) procedure nested in leave-pair-out cross validation (LPO-CV) is implemented during the SVM training. The SVM-based procedure has two main goals: to allow making a prediction of ASD diagnosis on unknown cases (subjects that have not entered the SVM training phase), with classification performance estimated in terms of either the accuracy or the area under the receiver operating characteristic curve (AUC); to directly extract a map encoding the anatomical differences between the patients with ASD and the matched control subjects (discrimination maps). The discrimination maps are shown in the figure at the beginning (e) and at the end (f) of the SVM-RFE procedure