Deep multiview learning to identify imaging-driven subtypes in mild cognitive impairment.

Area of Science:

• Neuroimaging analysis
• Alzheimer's Disease research
• Machine learning for medical data

Background:

• Multimodal imaging analysis in Alzheimer's Diseases (AD) research provides complementary information.
• Unsupervised clustering can identify disease subtypes but struggles with high-dimensional data.
• Existing methods applied directly to features may miss complex patterns in multimodal data.

Purpose of the Study:

• To identify multimodal imaging-driven subtypes in Mild Cognitive Impairment (MCI) participants.
• To utilize a multiview learning framework based on Deep Generalized Canonical Correlation Analysis (DGCCA).
• To learn shared, low-dimensional latent representations from three neuroimaging modalities.

Main Methods:

• Applied Deep Generalized Canonical Correlation Analysis (DGCCA), a multiview learning framework.
• Used neural networks for non-linear transformations to learn correlated, low-dimensional embeddings.
• Compared DGCCA embeddings with single modality features and linear Generalized Canonical Correlation Analysis (GCCA) using unsupervised clustering.

Main Results:

• DGCCA learned low-dimensional embeddings capturing more variance than GCCA.
• Amyloid PET imaging showed the most discriminative features within DGCCA.
• DGCCA-derived MCI subtypes exhibited differential cognitive, brain volume, and AD conversion patterns, confirming AD genetic markers.

Conclusions:

• DGCCA effectively learns low-dimensional embeddings from multimodal data via non-linear projections.
• MCI subtypes identified by DGCCA differ from early/late MCI classifications and align with amyloid PET findings.
• DGCCA subtypes demonstrate distinct cognitive and brain volume patterns, highlighting their potential for revealing disease structures.