Linking Brain Networks to Behavior with Subsampled Prediction and Stability Metrics

• Activity Number: 65
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: SSC
• Author(s): Stephen Strother*+
• Companies: Rortman Research Institute, Baycrest
• Keywords: fMRI ; prediction ; spatial stability ; regularization ; behavior ; brain networks

Abstract

An important goal of predictive multivariate models of fMRI data is discovering the structure of the spatial brain networks supporting experimental tasks and their behavioral measures. Such data sets have (no. brain locations) >> (no. observations) making model stability and overfitting critical concerns. Using a subsampling framework, "NPAIRS" I will examine tradeoffs between task prediction accuracy and the stability of estimated network patterns, for a canonical variates analysis regularized with a hierarchical PCA decomposition. In a multi-task fMRI study such prediction-stability tradeoffs revealed spatial network patterns strongly linked to task behavior. However, these patterns, while spatially highly stable, are only weakly to not predictive of the fMRI scans for particular tasks. Therefore, optimizing cross-validated prediction accuracy in so-called fMRI "mind-reading" analyses may not support the goal of extracting brain networks linked to behaviour; but by introducing spatial stability metrics such behavioral links can be revealed. Such results suggest the brain is spatially stable and temporally unstable, and our statistical approaches should reflect this tradeoff.