Classification of multiple sclerosis patients based on structural disconnection: A robust feature selection approach

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Diagnostics

Background:

• Structural disconnection is key in multiple sclerosis (MS) pathophysiology.
• Previous classification attempts using structural connectivity had low accuracy.

Purpose of the Study:

• To improve classification accuracy for MS patients using microstructure-informed tractography.
• To identify robust features from quantitative connectomes for MS classification.

Main Methods:

• Generated quantitative connectomes from diffusion MRI data of 55 MS patients and 24 controls.
• Employed a robust feature selection approach on network representations (whole connectivity, node strength, local efficiency).
• Tested classification accuracy with five classifiers and correlated features with clinical scales, comparing against standard thresholding methods.

Main Results:

• Identified 11 features for the whole network, 5 for local efficiency, and 7 for node strength.
• Achieved classification accuracies ranging from 64.5% to 91.1%, with whole network features performing best.
• Demonstrated superior performance compared to standard thresholding methods.
• Found correlations between selected features and clinical scales across motor, cognitive, fatigue, and depression domains.

Conclusions:

• A robust feature selection procedure applied to quantitative structural connectomes enables accurate MS patient classification.
• This method provides insights into white matter connections and gray matter regions most affected by MS pathology.