Improving fMRI-Based Autism Severity Identification via Brain Network Distance and Adaptive Label Distribution...

Area of Science:

• Neuroscience
• Machine Learning
• Developmental Psychology

Background:

• Autism spectrum disorder (ASD) diagnosis and severity assessment remain challenging due to label ambiguity and individual variability.
• Current functional magnetic resonance imaging (fMRI)-based methods for ASD severity identification lack satisfactory performance.
• The relationship between brain functional networks (BFN) and ASD symptom severity requires further investigation.

Purpose of the Study:

• To develop an advanced machine learning framework for accurate autism spectrum disorder (ASD) severity identification.
• To address limitations in current fMRI-based ASD severity assessment methods.
• To explore the association between BFN characteristics and ASD symptom severity.

Main Methods:

• Proposed a low- and high-level BFN distance (HBFND) method to construct BFN reflecting ASD severity differences.
• Utilized a multi-task network to account for individual variations in ASD communication and social skills.
• Employed an adaptive label distribution (ALD) technique to train the model and prevent overfitting.

Main Results:

• The proposed HBFND-AMLD framework demonstrated superior performance in ASD severity identification compared to state-of-the-art methods.
• The HBFND method effectively measured differences between individuals with ASD and healthy controls (HC) in low- and high-order BFN.
• The ALD technique successfully prevented model overfitting, enhancing identification accuracy.

Conclusions:

• The developed HBFND-AMLD framework shows significant potential for practical clinical diagnosis of ASD severity.
• This approach offers improved identification performance by considering BFN and individual differences in ASD.
• Further research into BFN and ASD severity could lead to more refined diagnostic tools.