An interpretable generative multimodal neuroimaging-genomics framework for decoding Alzheimer's disease.

Area of Science:

• Neuroimaging
• Genetics
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) is the leading cause of dementia globally.
• Mild cognitive impairment (MCI) is a prodromal stage of AD, with patients either progressing to AD or remaining stable.
• Multimodal data integration and handling missing data are crucial for accurate AD and MCI prediction.

Purpose of the Study:

• To develop a multimodal deep learning (DL) framework for classifying AD patients versus healthy controls.
• To detect MCI converters using multimodal MRI and single nucleotide polymorphisms (SNPs).
• To impute missing multimodal data and enhance model interpretability.

Main Methods:

• A multimodal DL framework incorporating a generative module (cycle GANs) for missing data imputation.
• Application of explainable AI (XAI) for feature relevance extraction and model interpretability.
• Utilized structural and functional MRI data alongside SNPs for classification and prediction tasks.

Main Results:

• Achieved high accuracy in AD detection (0.926±0.02) and MCI conversion prediction (0.711±0.01).
• Identified gray matter modulations in AD-associated cortical and subcortical brain regions.
• Revealed impairments in sensory-motor and visual resting-state networks and linked genetic mutations to endocytosis, amyloid-beta, and cholesterol pathways.

Conclusions:

• The integrative and interpretable DL approach demonstrates strong performance for AD detection and MCI prediction.
• The framework provides valuable biological insights into AD pathogenesis.
• This approach offers a promising tool for early diagnosis and understanding of AD and MCI.