Uncovering Active Structural Subspaces Associated with Changes in Indicators for Alzheimer's Disease.

Abstract

We present a framework for identifying subspaces in the brain that are associated with changes in biological and cognitive indicators for a given disorder. By employing a method called active subspace learning (ASL) on structural MRI features from an Alzheimer's disease dataset, we identify subsets of regions that form co-varying subspaces in association with biological age and mini-mental state exam (MMSE) scores. Features generated by projecting structural MRI components onto these subspaces performed equally well on regression tasks when compared to non-transformed features as well as PCA-based transformations. Thus, without compromising on predictive performance, we present a way to extract sparse subspaces in the brain which are associated with a particular disorder but inferred only from the neuroimaging data along with relevant biological and cognitive test measures.Clinical relevance-This work provides a way to identify active structural subspaces in the brain, i.e. subsets of brain regions which collectively change the most, in association with changes in the indicators of a given disorder.