Learning Active Multimodal Subspaces in the Brain.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Identifying brain patterns linked to demographic, biological, and cognitive factors is crucial for understanding neurological conditions.
• Current methods may not fully capture the complex interplay between structural and functional brain changes and clinical indicators.

Purpose of the Study:

• To introduce a novel multimodal framework for identifying brain subspaces defined by collective structural and functional changes.
• To apply this framework to understand biological ageing and schizophrenia using active subspace learning (ASL).

Main Methods:

• Developed a multimodal active subspace learning (ASL) framework to identify covarying subspaces in brain structure and function.
• Applied the method to a dataset examining biological ageing and schizophrenia, incorporating demographic, biological, and cognitive indicators.
• Generated multimodal features by projecting MRI components onto identified active subspaces.

Main Results:

• Successfully identified latent brain representations (subspaces) associated with biological age in the study sample.
• Demonstrated distinct subspace patterns in schizophrenia compared to cognitively normal individuals.
• Multimodal features derived from ASL outperformed PCA-based transformations and matched non-transformed features in a learning task.

Conclusions:

• The multimodal ASL method effectively learns active brain subspaces linked to specific conditions and traits from neuroimaging data.
• This approach offers a novel way to create multimodal brain biomarkers by identifying subspaces that maximally covary with cognitive or biological traits.
• These biomarkers retain significant predictive information, holding clinical relevance for understanding brain conditions.