Deep Multiview Learning to Identify Population Structure with Multimodal Imaging

Area of Science:

• Neuroimaging
• Computational Biology
• Machine Learning

Background:

• Identifying population structure is crucial for understanding disease heterogeneity.
• Multimodal imaging data offers rich information but poses analytical challenges.
• Existing methods may not fully leverage the complex relationships within multimodal data.

Purpose of the Study:

• To develop an effective deep multiview learning framework for population structure identification.
• To utilize multimodal imaging data for enhanced insights into Alzheimer's disease cohorts.
• To improve upon conventional generalized canonical correlation analysis (GCCA) methods.

Main Methods:

• Proposed a deep generalized canonical correlation analysis (DGCCA) framework.
• Learned a shared latent representation from non-linearly mapped, maximally correlated components.
• Applied cluster analysis to the DGCCA-derived feature set for population structure identification.

Main Results:

• DGCCA captured significantly more variance compared to linear GCCA.
• Identified a promising population structure within an Alzheimer's disease cohort.
• DGCCA-based population structure demonstrated a stronger genetic basis than competing methods.

Conclusions:

• Deep multiview learning with DGCCA is effective for population structure identification using multimodal imaging.
• The learned shared representation enhances genetic association analyses in disease cohorts.
• This framework offers a powerful tool for dissecting disease heterogeneity.