Group and Individual Non-Gaussian Components Analysis

Thursday, November 8, 2018 - 10:40am - 11:10am
Lind 305
David S. Matteson (Cornell University)
Independent component analysis (ICA) is an unsupervised learning method popular in functional magnetic resonance imaging (fMRI). Group ICA has been used to identify biomarkers in neurological disorders including Autism spectrum disorder [1] and dementia [2]. However, current group ICA methods use a PCA step that may remove important information associated with low-variance non-Gaussian features. Linear non-Gaussian component analysis (LNGCA) enables dimension reduction and component estimation simultaneously in single-subject fMRI [3]. We present a group and individual LNGCA model to extract group components shared by more than one subject, as well as subject-specific
individual components unique to each subject. To determine the number of components, we modify a parametric bootstrap approach [4] to account for spatial dependence. In simulation studies, our estimated group components achieve much higher accuracy compared to group ICA methods [5]. Moreover, we recover the individual components for each subject with high accuracy. We also apply our method to an fMRI experiment with healthy young adults in the Human Connectome Project, where the group signals include resting-state networks, and individual components include artifacts unrelated to neuronal signal. The decomposition into group and individual components is a promising direction for feature detection in neurological disorders.