A family of interpretable multivariate models for regression and classification of whole-brain fMRI data

Abstract

Increasing interest in applying statistical learning methods to functional magnetic resonance imaging (fMRI) data has led to growing application of off -the-shelf classification and regression methods. These methods allow investigators to use standard software packages to accurately decode perceptual, cognitive, or behavioral states from distributed patterns of neural activity. However, findings from these models are generally difficult to interpret when fit to whole-brain fMRI data, as they suffer from coefficient instability, are sensitive to outliers, and typically return dense rather than parsimonious solutions. Here, we develop several variants of the the graph-constrained elastic net (GraphNet)-a fast, whole-brain regression and classification method developed for spatial and temporally correlated data that yields interpretable model parameters. GraphNet models yield interpretable solutions by incorporating sparse graph priors based on model smoothness or connectivity as well as a global sparsity inducing prior that automatically selects important variables. Because these methods can be fit to large data sets efficiently and can improve accuracy relative to volume-of-interest (VOI) methods, they allow investigators to take an unbiased approach to model fitting-avoiding selection biases associated with VOI analyses or the multiple comparison problems of roaming VOI methods. As fMRI data are unlikely to be Gaussian, we (1) extend GraphNet to include robust loss functions that confer insensitivity to outliers, (2) equip them with "adaptive" penalties that have oracle properties, and (3) develop a novel sparse structured support vector machine (SVM) GraphNet classifier to allow maximum-margin classification. When applied to previously published data, these efficient whole-brain methods significantly improved classification accuracy over previously reported VOI-based analyses.