Predicting Motor Tasks in fMRI Data with Support Vector Machines
Stephen M. LaConte, Stephen C. Strother, Vladimir Cherkassky, Xuting Hu · 2003
S. LaConte, S. Strother, V. Cherkassky, X. Hu Emory University/Georgia Tech, Atlanta, GA, United States, University of Minnesota, Minneapolis, MN, United States SYNOPSIS The support vector machine (SVM) is introduced to fMRI as a method for classifying temporal scans. The SVM is found to perform comparably to canonical variates analysis (CVA) in terms of misclassification error. We examine the interpretation of the SVM model in the context of fMRI, and find that removing model-related scans enhances the statistical difference between those scans. INTRODUCTION This study focuses on temporally predictive models of fMRI data. Specifically, we introduce the support vector machine (SVM), a universal supervised learning method arising out of the statistical learning theory of Vapnik [1,2]. Our focus on temporal prediction is motivated by three key observations: i) Temporal information about an fMRI experiment is critical for either directly obtaining summary maps, or indirectly interpreting “data-driven” results. ii) As recently demonstrated with CVA, classification of temporal scans can be used in model validation [3]. iii) The unique formulation of the SVM, which can find non-linear decision boundaries and handle high dimensional problems, makes it a promising technique for fMRI. We compare the classification accuracy of the SVM to that of canonical variates analysis (CVA) using cross-validation on data from a visually guided static force task. A key issue that needs to be addressed for the neuroscientific applicability of the SVM to fMRI is the physical meaning associated with a given support vector model. We begin to address this interpretation problem by looking at the SVM model of a simple bi-manual finger opposition task. THEORY Here we summarize only the salient concepts for SVM-based classification (see [3,4] for a description of CVA). Training data, consisting of input vectors,