A Denoising Variational Autoencoder for the Diagnosis of Autism based on Resting-state fMRI

Health Analytics
Written By Alexis Avedisian

Video

Team Information

Team Members

  • Xinyuan Zheng, Department of Statistics, Graduate School of Arts and Sciences, Columbia University

  • Faculty Advisor: Xiaofu He, Assistant Professor of Clinical Neurobiology, Vagelos College of Physicians and Surgeons

Abstract

Autism Spectrum Disorder (ASD) affects the perception, social interactions and communication skills of the patients and has complicated phenotypes. In this study, we used the public autism dataset from Paris-Saclay Center for Data Science, which consists the fMRI data collected from 549 ASD patients and 601 healthy controls, and apply machine learning methods to the diagnosis of ASD. We proposed a feature extraction model, a denoising variational autoencoder (DVAE), for the classification of the brain images for ASD. This method is tested on a hold-out test set and the classification accuracy of our best model is around 60% before and after feature reduction. Further work could potentially rectify the low accuracy by training separate models for controls and patients, taking the differences across sites into account and using larger datasets.

Team Lead Contact

Xinyuan Zheng: xz2906@columbia.edu

Alexis Avedisian
Previous

Ureteral Stent Decompression is Associated with Decreased Length of Stay in Obstructive, Infected Ureteral Stones
Next

iSamples (Internet of Samples): Cyberinfrastructure to Support Transdisciplinary Use of Material Samples