Multimodal subspace independent vector analysis effectively captures the latent relationships between brain structure...

Area of Science:

• Neuroscience
• Neuroimaging Analysis
• Biostatistics

Background:

• Understanding brain structure-function relationships from high-dimensional, multimodal neuroimaging data is a key neuroscience challenge.
• Conventional methods often oversimplify statistical assumptions, limiting the capture of complex, multi-dimensional relationships within and between brain data modalities.
• Existing approaches struggle to account for subject-level variability in latent brain sources.

Purpose of the Study:

• To introduce Multimodal Subspace Independent Vector Analysis (MSIVA), a novel methodology for analyzing high-dimensional, multimodal neuroimaging data.
• To capture joint and unique vector sources from multiple data modalities by defining flexible, variable-dimension subspaces.
• To enable the estimation of subject-level variability within independent subspaces, overcoming limitations of traditional methods.

Main Methods:

• Developed Multimodal Subspace Independent Vector Analysis (MSIVA) to define cross-modal and unimodal subspaces with variable dimensions.
• Enabled flexible estimation of independent subspaces within modalities and their linkage across modalities.
• Compared MSIVA against unimodal and multimodal baseline methods using synthetic and real neuroimaging datasets (sMRI, fMRI) with varying subspace structures.

Main Results:

• MSIVA successfully identified ground-truth subspace structures in synthetic datasets, outperforming a multimodal baseline that failed to detect high-dimensional subspaces.
• MSIVA demonstrated superior detection of latent subspace structures in large multimodal neuroimaging datasets (sMRI/fMRI) compared to a unimodal baseline.
• Subspace-specific analyses revealed strong associations between MSIVA-derived sources and phenotype variables (age, sex, schizophrenia, lifestyle, cognition).

Conclusions:

• MSIVA effectively captures complex, multi-dimensional relationships in multimodal neuroimaging data, including subject-level variability.
• The method identified modality- and group-specific brain regions associated with key phenotype measures, highlighting potential biomarkers for neurological and psychiatric conditions.
• Findings suggest MSIVA provides valuable insights into the linked brain structure and function underlying various phenotypic traits and neuropsychiatric disorders.