International Meeting for Autism Research: Applying Machine Learning Techniques to Brain Imaging Characteristics to Distinguish Between Individuals with Autism and Neurotypical Controls

Applying Machine Learning Techniques to Brain Imaging Characteristics to Distinguish Between Individuals with Autism and Neurotypical Controls

Friday, May 21, 2010
Franklin Hall B Level 4 (Philadelphia Marriott Downtown)
9:00 AM
S. E. Schipul , Psychology, Center for Cognitive Brain Imaging, Carnegie Mellon University, Pittsburgh, PA
S. Aryal , Psychology, Center for Cognitive Brain Imaging, Carnegie Mellon University, Pittsburgh, PA
M. A. Just , Psychology, Center for Cognitive Brain Imaging, Carnegie Mellon University, Pittsburgh, PA
Background: Autism spectrum disorder is a genetic neurodevelopmental disorder characterized by deficits in language, social interaction, and repetitive behaviors. Although many neuroimaging studies have shown underlying differences in the brain structure and function of individuals with autism as compared to neurotypical controls, currently the only method of diagnosing autism is through interviews and evaluations with trained clinicians.  Previous studies (Ecker et al., 2009; Fahmi et al., 2007) have explored the potential of structural brain measures to predict a diagnosis of autism. However, brain activation and synchronization may also be able to contribute to such predictions.

Objectives: This project investigated the potential of machine learning algorithms to distinguish between individuals with autism and neurotypical individuals based on their brain structure, activation, and synchronization.

Methods: fMRI and MRI data were collected for 43 high-functioning individuals with autism and 43 neurotypical control participants. During the fMRI scan, participants read sentences and decided if they were true or false, interspersed with a fixation condition.  The data submitted to the machine learning algorithms included activated voxel counts within several regions of interest (ROIs) during fixation, functional connectivity (synchronization) measures between pairs of ROIs obtained during the task performance, and white matter measurements from MRI scans. The classification algorithms that were used to predict a diagnosis of autism included Gaussian Naïve Bayes, logistic regression, and support vector machines.

Results: Our classifiers were able to distinguish between individuals with autism and neurotypical controls with an accuracy of 66%, where the group membership was established using ADOS scores and expert clinical diagnosis.

Conclusions: These findings suggest that brain imaging data has potential to play a role in diagnosing autism. Using several different machine learning algorithms, we were able to distinguish between individuals with autism and neurotypical controls with an accuracy well above chance based on brain imaging data concerning structure, activation, and synchronization.

See more of: Brain Imaging
See more of: Brain Imaging
See more of: Brain Structure & Function