Structural MRI Does Not Support the DSM-5 Unification of the DSM-IV-TR Autism Spectrum Diagnoses

Poster Presentation
Friday, May 11, 2018: 11:30 AM-1:30 PM
Hall Grote Zaal (de Doelen ICC Rotterdam)
E. Ferrari1, A. Giuliano1, P. Bosco1, P. Oliva2, M. E. Fantacci3, F. Muratori4, S. Calderoni5 and A. Retico6, (1)National Institute for Nuclear Physics (INFN), Pisa, Italy, (2)University of Sassari and INFN, Physics Department, Sassari, Italy, (3)University of Pisa, Physics Department, Pisa, Italy, (4)IRCCS Stella Maris Foundation, Calambrone (Pisa), Italy, (5)Department of Clinical and Experimental Medicine, University of Pisa, Pisa, Italy, (6)Pisa Division, National Institute for Nuclear Physics, Pisa, Italy
Background: Autism Spectrum Disorders (ASD) are a heterogeneous condition that affects individuals with various degrees of severity. Brain magnetic resonance imaging (MRI) represents a valuable non-invasive technique to study this condition. In literature, there is a great amount of studies based on supervised classification algorithms in order to distinguish subjects with ASD from controls through an analysis of their structural brain images. However, the results obtained are controversial and usually not reproduced on a sufficiently wide statistical sample. Recently, many studies focused on the importance of the stratification of the ASD population to improve classification performance.

Objectives: To evaluate the impact of stratification on supervised classification of subjects with Autism Spectrum Disorders. To this purpose both the diagnostic categories of the DSM-IV-TR and DSM-5 have been taken into account.

Methods: In this study, 420 morphological features extracted with Freesurfer 6.0 from 2156 subjects from the public databases ABIDE I and ABIDE II [http://fcon_1000.projects.nitrc.org/indi/abide/] were analyzed. In first place, an outlier analysis has been conducted to exclude the subjects for which the brain segmentation was unsuccessful. Then, six different supervised machine-learning algorithms have been used on different groups of subjects to study the effects of stratification; this has been repeated using different feature normalization methods. For each classifier the area under the receiver operating characteristic (AUC) has been computed. Finally, an analysis of the most relevant features for the decision-making process has been conducted.

Results: The great amount of train-test experiments we performed allowed us to obtain the following main results: 1) the feature normalization does not have a significant effect on the AUCs; 2) by contrast, the stratification and the type of classification algorithm used have a strong impact. Specifically, using a Logistic regression classifier, whilst the entire group of ASD subjects (according to DSM-5 classification) differed from controls with a modest AUC (about 0.60), once the subjects are sub-grouped according to the DSM-IV-TR separate subcategories of autism (AD), Asperger (AS) and Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS), the AUC values achieved remain limited for AD vs. controls, AD vs. AS and AS vs. controls, whereas they are: 0.82 for AS vs. PDD-NOS, 0.80 for AD vs. PDD-NOS and 0.82 for PDD-NOS vs. controls.

Conclusions: We found out that the stratification of the ASD population in more homogeneous subgroups according to the DSM-IV-TR diagnoses leads to an improvement of the case vs. control classification performance. Despite the AUC values we obtained are still far away from allowing us to claim for the discovery of a neuroimaging-based biomarker for ASD, the identification of which brain regions are more responsible for the subgroup separations may provide new clues in the understanding of the neurobiology underling the ASD condition.