Exploring Implicit Biological Heterogeneity in ASD Diagnosis Using a Multi-Head Attention Graph Neural Network

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Autism spectrum disorder (ASD) presents heterogeneous characteristics, influenced by sex and age, with current deep learning models for functional connectivity (FC) graphs lacking specialized regional analysis.
• Existing methods focusing on generalized patterns fail to capture intricate, variable brain connectivity crucial for accurate ASD diagnosis.

Purpose of the Study:

• To develop a novel deep learning method for modeling FC with multi-head attention to overcome limitations in analyzing ASD-related brain connectivity.
• To accurately assess disease indications by extracting abnormal patterns in brain connectivity, considering region-specific correlations and transient time points.

Main Methods:

• Proposed a deep learning method that models functional connectivity (FC) using multi-head attention to capture intricate and variable patterns.
• Transformed FC data into a graph with weighted edge labels, processed by a graph neural network capable of handling edge labels.
• Utilized the Autism Brain Imaging Data Exchange (ABIDE) I and II datasets for model validation.

Main Results:

• The novel method demonstrated superior performance over state-of-the-art techniques on the ABIDE datasets, improving diagnostic accuracy by up to 3.7%p.
• Multi-head attention significantly enhanced the differentiation between typical and ASD brains by analyzing FC.
• Ablation studies confirmed the method's ability to validate diverse brain characteristics across different ages and sexes in ASD patients.

Conclusions:

• The proposed deep learning method effectively enhances diagnostic accuracy for ASD.
• This approach holds significant potential for advancing neurological research and improving ASD diagnosis.