Different neuroimaging approaches have been proposed to date to explore the genetic, clinical and neurobiological heterogeneity of Autism Spectrum Disorders (ASD).¹ We propose the implementation of One-Class Classifiers (OCC) based on Support Vector Machines (OCC-SVM) to study the heterogeneous neuroanatomical profile of ASD, based on MRI data. In contrast to conventional binary classification algorithms, the OCC or Data Description methods² make a description of the positive sample only (referred to also as the target class) and test the similarity of new cases to this target class, eventually classifying them as outliers. Taking inspiration from the two-class SVM, Tax and Duin addressed the OCC problem by proposing a method to obtain a spherically shaped boundary around the target sample.³ We implemented the OCC-SVM to analyze Freesurfer-based regional characteristics extracted from structural MRI to measure their performance in the discrimination of young subjects with ASD from controls. We used the OCC-SVM to investigate whether the distribution of normal patterns of brain structure features is enough homogeneous within the control population to enable the definition of a robust boundary in relation to which the ASD patients are classified as outliers. As an alternative, a consistent pattern among the patients with ASD could provide a boundary according to which the controls are classified as outliers.A population of 81 children in the age range of 24-72 months (21 males with ASD, 20 male controls, 20 females with ASD and 20 female controls) was considered. Patients and controls were accurately matched for age and non-verbal IQ (see Table 1). Structural MRI data were acquired at 1.5 T (FSPGR, TR=12.4ms, TE=2.4ms, TI=700ms, FA=10°, slice thickness=1.1mm, in-plane resolution=1.1x1.1mm²) and processed according to the Freesurfer analysis pipeline (http://surfer.nmr.mgh.harvard.edu/) to obtain five descriptive features (Surface Area, Volume, Cortical Thickness, Standard Deviation of Cortical Thickness, Mean Curvature) for each of the 62 cortical structures of the Desikan-Killiany-Tourville atlas.⁴ To train and test the OCC we used RapidMiner (http://rapidminer.com/) advanced analytics platform version 5.3, which includes the OCC-SVM as a part of the LibSVM operator. The optimization of the training parameters (ν and γ for the OCC-SVM with radial basis function kernel) and the evaluation of the classification performance, in terms of the area under the receiver operating characteristic curve (AUC), were conducted according to a nested leave-pair-out cross-validation (LPO-CV) protocol. We implemented the algorithm proposed by Schölkopf et al.⁵ to generate a preimage representing the neuroanatomical regions most contributing to the OCC boundary definition. We adopted the permutation testing procedure to assign a statistical significance to the regions we identified,⁶ after tailoring it to the OCC method.We first considered the control subjects as the target class, and we found out that the hypersphere (or decision boundary) enclosing most of the controls contained also most ASD data points. The separation ability of the OCC-SVM fell in this case within the chance level. By contrast, when we considered the ASD group as the target class, the performance achieved in the ASD vs. control separation, evaluated according to the LPO-CV protocol, were: AUC=0.74 for the male subset, AUC=0.68 for the female subset and AUC=0.64 for the entire dataset. It means that the OCC-SVM could capture a common structure among the brain MRI features of the ASD subjects. Finally, we identified with the preimage method and the permutation testing with 10000 iterations (p<0.05) the brain regions most relevant to the definition of the OCC decisional boundary, as reported in Figure 1. They belong to a network of structural brain alterations widely reported in the population with ASD, including frontal and temporal grey matter areas for both the male and female populations.OCC are a suitable method to analyze data when the positive class is well characterized, whereas the negative class is not sufficiently representative of the negative population, as it happens in our analysis. Despite the accurate case-control matching for age, gender and non-verbal IQ performed in our data sample, we found out evidence of a higher heterogeneity in the brain features of the control group with respect to the ASD patients’ ones. This hampers the OCC-SVM trained on the control samples achieving good classification performance in case-control separation. By contrast, the ASD groups showed a common structure of features that the OCC-SVM could capture.The OCC-SVM identified consistent patterns of alterations in neuroimaging data of the ASD population. Despite the phenotypic heterogeneity of ASD, a common neuroanatomical profile involving frontal and temporal grey matter areas underlying the core features was detected.