A network clustering based feature selection strategy for classifying autism spectrum disorder.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Non-invasive neuroimaging techniques like resting-state functional magnetic resonance imaging (rs-fMRI) are crucial for studying brain function and structure.
• Whole-brain functional networks derived from rs-fMRI are commonly used to investigate neurological conditions such as autism spectrum disorder (ASD).
• Current ASD classification methods relying on whole-brain network features may lack sufficient discriminative power.

Purpose of the Study:

• To develop an improved feature selection strategy for accurate ASD classification.
• To investigate the efficacy of using subnetwork features for ASD diagnosis.

Main Methods:

• A network clustering-based feature selection strategy was proposed.
• Symmetric non-negative matrix factorization was employed to partition brain networks into four modules.
• Features were extracted specifically from the default mode network (DMN) module for classifier training.

Main Results:

• The proposed method, utilizing features from the DMN, demonstrated superior performance compared to methods using whole-brain network features.
• Computational experiments validated the effectiveness of the DMN-based feature extraction approach.

Conclusions:

• Feature selection based on the default mode subnetwork is a promising strategy for enhancing ASD classification.
• This approach offers a more discriminative basis for developing accurate ASD diagnostic tools.