EIGNN: An Explainable Imaging-Genetic Neural Network for Robust Alzheimer's Disease Risk Prediction.

Area of Science:

• Neuroscience
• Genetics
• Artificial Intelligence

Background:

• Accurate Alzheimer's disease (AD) risk prediction and early diagnosis are vital for timely intervention.
• Existing models struggle with complex data interactions and lack biological explainability for clinical use and biomarker discovery.

Purpose of the Study:

• To develop an Explainable Imaging-Genetic Neural Network (EIGNN) for accurate, robust, and explainable AD risk prediction.
• To integrate genetic and neuroimaging data using a biologically informed approach.

Main Methods:

• Developed a biologically-informed neural network architecture (EIGNN).
• Incorporated multi-GWAS SNP selection, enhanced explainable neural network design, and modal attention.
• Hierarchically mapped genetic variants to genes and pathways, integrating them with neuroimaging features.

Main Results:

• The EIGNN model demonstrated superior performance compared to existing methods.
• Achieved improved robustness, explainability, and reproducibility in AD risk prediction.
• Identified potential AD risk genes and pathways through multi-modal feature interaction mapping.

Conclusions:

• The EIGNN model offers a powerful tool for AD risk prediction by effectively integrating multi-modal data.
• The model's explainability facilitates clinical adoption and aids in discovering novel AD biomarkers.
• Further exploration of gene-brain region interactions can deepen our understanding of AD pathogenesis.