Distributed Intrinsic Functional Connectivity Patterns Predict Diagnostic Status in Large Autism Cohort.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Imaging

Background:

• Autism spectrum disorder (ASD) diagnosis relies on behavioral observation due to lack of established brain markers.
• Previous machine learning studies using functional connectivity MRI (fcMRI) reported modest diagnostic accuracies (60-80%).
• Feature correlation in fcMRI data can bias traditional machine learning models.

Purpose of the Study:

• To develop a more accurate machine learning model for ASD diagnosis using fcMRI data.
• To investigate the utility of conditional random forest (CRF) for handling feature correlation in fcMRI.
• To identify informative brain connectivity patterns associated with ASD.

Main Methods:

• Utilized resting-state functional MRI (rs-fMRI) data from 126 ASD and 126 typically developing (TD) participants from the Autism Brain Imaging Data Exchange (ABIDE).
• Employed a conditional random forest (CRF) classifier, an ensemble method robust to feature correlation.
• Classified participants based on functional connectivity matrices derived from 220 brain regions of interest.

Main Results:

• CRF models achieved high classification accuracies of 92-99% with over 300 informative features.
• Pericentral somatosensory and motor regions showed disproportionate informativeness.
• External validation on an independent dataset yielded lower accuracies (67-71%), likely due to sample heterogeneity.

Conclusions:

• Conditional random forest models demonstrate significant potential for accurate ASD diagnosis from fcMRI data.
• The high accuracy suggests optimized characterization of the studied cohort, not necessarily overfitting.
• Etiological heterogeneity within ASD may explain the large number of informative features and variations in external validation accuracy.