Functional connectivity-based classification of autism and control using SVM-RFECV on rs-fMRI data

Area of Science:

• Neuroscience
• Computational Psychiatry
• Biomedical Engineering

Background:

• Autism spectrum disorder (ASD) classification accuracy is limited by current feature selection methods.
• Identifying reliable biomarkers for ASD remains a significant challenge in neuroscience.

Purpose of the Study:

• To develop and validate a functional connectivity (FC)-based algorithm for improved ASD classification.
• To identify optimal features for distinguishing individuals with ASD from neurotypical controls using machine learning.

Main Methods:

• Utilized a support vector machine-recursive feature elimination (SVM-RFE) algorithm on a functional connectivity (FC) matrix derived from 35 regions of interest.
• Employed a stratified 4-fold cross-validation strategy for feature selection and a Gaussian kernel SVM for classification.
• Analyzed data from 255 individuals with ASD and 276 neurotypical controls across 10 different research sites.

Main Results:

• The proposed algorithm achieved a global classification accuracy of 90.60% (sensitivity 90.62%, specificity 90.58%).
• Cross-site validation demonstrated robust performance, with classification accuracy ranging from 75.00% to 95.23% on the leave-one-site-out test.
• The SVM-RFE method effectively identified a discriminative subset of FC features.

Conclusions:

• The developed FC-based algorithm significantly improves ASD classification accuracy compared to previous studies.
• The approach accurately measures feature importance, selecting only the most discriminative functional connectivity patterns.
• This method holds promise for more precise and reliable diagnosis of autism spectrum disorder.