Identification of Autism Based on SVM-RFE and Stacked Sparse Auto-Encoder

Abstract

In order to improve the classification accuracy of patients with autism based on the full Autism Brain Imaging Data Exchange dataset, a total of 501 subjects with autism and 553 subjects with typical control across 17 sites were involved in the study. Firstly, we applied the resting-state functional magnetic resonance imaging data to calculate the functional connectivity (FC) based on the automated anatomical labeling atlas with 116 brain regions. Secondly, we adopted the support vector machine-recursive feature elimination algorithm to select top 1000 features from the primitive FC features. Thirdly, we trained a stacked sparse auto-encoder with two hidden layers to extract the high-level latent and complicated features from the 1000 features. Finally, the optimal features obtained were fed into the softmax classifier. Experimental results demonstrate that the proposed classification algorithm is able to identify the autism with a state-of-the-art accuracy of 93.59% (sensitivity 92.52%, specificity 94.56%).

Keywords

• Autism
• fMRI
• deep learning
• SVM-RFE
• classification