Flexible Multimodal Neuroimaging Fusion for Alzheimer's Disease Progression Prediction

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder characterized by variable rates of cognitive decline.
• Accurate prediction of AD progression is crucial for patient care and clinical trial design.
• Multimodal neuroimaging data integration offers potential for improved prediction, but existing models struggle with missing data during inference.

Purpose of the Study:

• To develop a flexible multimodal model for Alzheimer's disease progression prediction that performs well even with significant missing neuroimaging data.
• To introduce PerM-MoE, a novel sparse mixture-of-experts approach with independent modality routers.
• To compare the performance of PerM-MoE against state-of-the-art models and unimodal approaches in predicting cognitive decline.

Main Methods:

• Utilized T1-weighted MRI, FLAIR, amyloid beta PET, and tau PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort.
• Developed PerM-MoE, a sparse mixture-of-experts model featuring independent routers for each neuroimaging modality.
• Evaluated model performance in predicting two-year changes in Clinical Dementia Rating-Sum of Boxes (CDR-SB) scores under various missing data scenarios, comparing against Flex-MoE and unimodal models.

Main Results:

• PerM-MoE demonstrated superior performance in predicting Alzheimer's disease progression (CDR-SB change) across most tested levels of missing neuroimaging data.
• The proposed PerM-MoE model showed more effective utilization of available experts compared to the state-of-the-art Flex-MoE.
• The independent router mechanism in PerM-MoE enhances flexibility and predictive power in data-scarce situations.

Conclusions:

• PerM-MoE offers a robust and flexible solution for Alzheimer's disease progression prediction using multimodal neuroimaging, particularly excelling in clinically relevant scenarios with missing data.
• The independent router design is key to PerM-MoE's improved performance and adaptability.
• This approach advances the utility of multimodal data for forecasting cognitive decline in Alzheimer's disease.