Personalized Explanations for Early Diagnosis of Alzheimer's Disease Using Explainable Graph Neural Networks with...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Data Science

Background:

• Alzheimer's disease (AD) diagnosis and prognosis remain challenging.
• Graph neural networks (GNNs) offer novel approaches for analyzing complex biological data.
• Understanding AD progression requires integrating diverse patient data.

Purpose of the Study:

• To apply graph convolutional networks (GCNs) to a correlation-based population graph for improved AD prognosis.
• To predict amyloid-beta (Aβ) positivity using demographic and neuroimaging data.
• To identify key biomarkers associated with Aβ positivity and AD heterogeneity.

Main Methods:

• Development and application of a GCN model utilizing a correlation-based population graph.
• Comparison of the GCN model against conventional machine learning and a baseline GCN with random edges.
• Utilizing GNNExplainer for model interpretability and biomarker identification.

Main Results:

• The correlation-based GCN model demonstrated superior performance in predicting Aβ positivity compared to other methods.
• The model showed consistent effectiveness across different patient groups within the ADNI dataset.
• GNNExplainer identified distinct sets of biomarkers associated with Aβ positivity, highlighting AD heterogeneity.

Conclusions:

• The proposed GCN approach enhances AD prognosis by accurately reflecting correlation structures in population graphs.
• This method offers potential for more personalized therapeutic strategies in AD.
• The approach has broad applicability for complex disease research and personalized medicine.