A Hierarchy of Brain Networks Revealed by MVPA Performance Metrics

• Activity Number: 144 - Uncover the Essential Truth by Integrating Big and Complex Imaging Data with New Statistical Tools
• Type: Invited
• Date/Time: Monday, July 31, 2017 : 10:30 AM to 12:20 PM
• Sponsor: SSC
• Author(s): Stephen Strother* and Cheryl Grady
• Companies: Baycrest & University of Toronto and Baycrest & University of Toronto
• Keywords: BOLD fMRI ; brain networks ; multivariate pattern analysis ; regularization ; spatial reproducibility ; prediction

Abstract

BOLD fMRI evidence for task-dependent, network hierarchies in the brain appear in metric plots designed to compare the performance of multivariate pattern analysis (MVPA) models. These pseudo-ROC performance plots are defined by subsampled prediction (P) plotted against spatial network reproducibility (R) metrics. (P, R) plots of linear discriminant (LD) models regularized with dimensionality reduction using principal components (PC) reveal task-dependent hierarchies in the brain's network structure. During task performance an LD model regularized with PCs added from largest to smaller variance traces out a characteristic (P, R) curve shape as a function of the # PCs (q). Such (P, R) curves typically start with Pmin equivalent to random guessing for q=1, and rise to a task-dependent Pmax for q [20,100]. Values of R typically start near Rmax for q=1, decline and reach a local maximum before Pmax, and then decrease rapidly with increasing q. I will present evidence that the resulting normalized, spatial patterns as a function of increasing P and q reflect the regional hierarchies of underlying brain-network continua adapted to meet particular task demands.