Multimodal subspace independent vector analysis captures latent subspace structures in large multimodal neuroimaging studies.

We present Multimodal Subspace Independent Vector Analysis (MSIVA), a methodology to capture both joint and unique vector sources across multiple data modalities by defining linked and modality-specific subspaces. In particular, MSIVA enables the estimation of independent subspaces of various sizes within modalities and their one-to-one linkage to corresponding subspaces across modalities. We compare MSIVA to a fully unimodal initialization baseline and a fully multimodal initialization baseline, and evaluate all three approaches with five distinct subspace structures on synthetic and neuroimaging datasets. We first demonstrate that MSIVA and the unimodal baseline can identify the correct ground-truth subspace structures from the incorrect ones in multiple synthetic datasets, while the multimodal baseline fails at detecting high-dimensional subspace structures. We then show that MSIVA can better capture the latent subspace structure with the minimum loss value in two large multimodal neuroimaging datasets compared to the unimodal baseline. Our results from subsequent per-subspace canonical correlation analysis (CCA) and brain-phenotype modeling demonstrate that the sources from the optimal subspace structure are strongly associated with phenotype measures, including age, sex and schizophrenia-related effects. Our proposed methodology MSIVA can be applied to capture linked and unique biomarkers from multimodal neuroimaging data.