Autism Diagnosis Using Iterative Permutation Sampling-Recursive Feature Elimination Algorithm and Deep Learning

Autism spectrum disorder (ASD) is a neuro-developmental disorder that affects social and communication abilities. There are no confirmed causative factors for the spectrum of symptoms that occur in ASD children. Currently, the gold standard for an ASD diagnosis is based on clinical testing. In particular, brain imaging modalities are believed to hold discriminant information for an ASD diagnosis. Recently, it has been proposed that altered functional connectivity patterns in the resting state functional MRI (RfMRI) coupled with machine learning may hold promise for an ASD diagnosis. However, algorithms that extract these patterns generate a large number of connectivity features, leading to high dimensional data. To address this problem, we propose a novel efficient feature selection algorithm called Iterative Permutation Sampling-Recursive Feature Elimination (IPS-RFE). Only a limited number of infor-mative discriminating features are fed to a deep neural network classifier. We have investigated this approach for classifying ASD in the ABIDE 1 dataset which contains approximately 1000 subjects. The proposed feature selection and classification approach outperforms other state-of-the-art alternatives with an accuracy of 75%, sensitivity of 73.5%, specificity of 76.5% and area under ROC curve of 0.803. A high percentage of the features selected by the IPS- RFE algorithm belong to the default mode, limbic, and visual brain networks, which have been reported to be abnormal among ASD children.