A joint network optimization framework to predict clinical severity from resting state functional MRI data.

Area of Science:

• Neuroimaging
• Machine Learning
• Clinical Neuroscience

Background:

• Resting-state fMRI (rs-fMRI) is crucial for understanding brain function in neurological disorders.
• Predicting clinical severity from rs-fMRI data remains challenging due to complex brain network interactions.
• Existing methods often fail to capture both group-level effects and individual patient heterogeneity.

Purpose of the Study:

• To develop a novel optimization framework for predicting clinical severity from rs-fMRI data.
• To identify robust biomarkers for Autism Spectrum Disorder (ASD) using brain network analysis.
• To improve upon existing semi-supervised learning techniques in neuroimaging.

Main Methods:

• A two-term coupled optimization framework was developed.
• The first term decomposes correlation matrices into sparse subnetworks using rank-one outer-products.
• The second term uses patient-specific coefficients in a linear regression model to predict clinical severity.
• Validation was performed on two independent ASD cohorts using ten-fold cross-validation.

Main Results:

• The proposed framework significantly outperforms standard semi-supervised methods.
• It successfully captures both group-level effects and patient heterogeneity in rs-fMRI data.
• The method robustly identifies clinically relevant brain networks associated with ASD.

Conclusions:

• The novel joint network optimization framework offers a powerful approach for predicting clinical severity from rs-fMRI.
• This method enhances our understanding of brain network alterations in ASD.
• It provides a robust tool for identifying ASD-specific neural signatures.