A Learning Framework For Multimodal Active Subspaces In The Brain

Multimodal neuroimaging data are complex and high-dimensional, and the multimodal inter-relationships between brain regions can be hard to estimate and interpret. To address this, we present a support vector regression-based active subspace learning (SVR-ASL) framework to learn sparse brain subspaces that span both structural and functional modalities, and collectively co-vary maximally in association with a given cognitive or biological trait. We apply this approach to a multimodal schizophrenia (SZ) dataset. We show that, unlike similar PCA-based transformations, these sparse subspaces successfully encode associations with the studied trait (age) while retaining predictive accuracy in an informative manner.