Multimodal Subspace Independent Vector Analysis Better Captures Hidden Relationships In Multimodal Neuroimaging Data

We introduce multimodal subspace independent vector analysis (MSIVA), a methodology to capture both joint and unique latent sources across data modalities by defining shared and modality-specific subspaces. We compared MSIVA to a unimodal analysis (UA) baseline and tested both methods with four distinct subspace structures on synthetic and multimodal neuroimaging datasets. We demonstrated that both approaches can identify and distinguish the correct subspace structures from incorrect ones on synthetic datasets. We then showed that MSIVA can better capture the subspace structures across two neuroimaging modalities, compared to UA, and output the optimal subspace structure with the minimum loss on the neuroimaging dataset. Results from subsequent per-subspace canonical correlation analysis and brain-phenotype modeling showed that the sources from the optimal subspace structure are significantly associated with phenotype measures including age and sex.