Abstract
Introduction
Pattern recognition techniques promise improved sensitivity and flexibility for the analysis of functional MRI (fMRI) data (Haynes and Rees 2006). This work presents a one-step approach to a group analysis (SVM-GA) which is based on the use of a support vector machine (Vapnik 1995). It compares SVM-GA performance – and the information it provides – with that of a conventional two-step group analysis by generalized linear model (GLM).
Methods
SVM-GA and GLM were applied to fMRI data of a bimanual sequential finger-opposition task (Baraldi, Porro et al. 1999) using both the full data and a reduced data set with and without spatial filtering.
Activation maps from the group of subjects were estimated within a regular two-level GLM framework. A first-level single-subject analysis with a canonical HRF was followed by a higher-level analysis applying mixed-effect (ME) models (using FLAME1 = FMRIB Local Analysis for Mixed Effects) and fixed-effect (FE) models in FEAT. Activation maps were thresholded either by controlling False Discovery Rate (Genovese, Lazar et al. 2002) (FDR, no need of spatial filtering) or by employing random field theory for multiple-testing adjusted significance (spatial filtering required), using clusters determined by Z > 2.3 and a corrected cluster significance threshold of p = 0.05 (Hayasaka, Phan et al. 2004).
SVM-GA directly calculated spatial discrimination maps for group data: SVM input was the group signal at the voxel, with repeated measurements as sample units, and the condition (active or passive) as classification target. Condition was labeled according to the experimental block design timing, with a shift of 6 s to account for hemodynamic delay. SVM leave-one-out generalization performance was evaluated directly from the trained SVM by generalized XiAlpha estimates (Joachims 2000). The latter contain an appropriate bias correction and therefore avoid the need of repeated SVM training for leave-one-out cross-validation. Spatial maps of SVM generalization performance were converted into spatial maps of statistical p-value based on a permutation approach. The latter simulated the distribution of the generalized XiAlpha estimate under the statistical null-hypothesis of no true SVM generalization at an arbitrary voxel. Maps of statistical p-value were thresholded for multiple-testing adjusted significance by established procedures such as, e.g., False Discovery Rate (Genovese, Lazar et al. 2002), iterative-Two-Thresholds (Auer, Schweizer et al. 2011), or by a SVM-specific filter method. Voxels with significant SVM generalization were considered to be informative, i.e., involved in the task. SVM weights quantify individual subject contributions to the classifier.
Results
SVM-GA successfully recognized discriminating patterns in the primary somatomotor cortex and supplementary motor area similar to GLM. In contrast to GLM, SVM-GA had increased power (Figure 1) attributed to its robustness with respect to imperfections of spatial registration and between-subject variance in BOLD delay. This was demonstrated by revealing more clearly patterns in the primary somatosensory cortex (S1), which are in line with a bimanual task. SVM-GA – at the same time – also provided additional subject-level reports (Figure 2), which may explain the results and reveal outlying subjects. It also allowed simple estimation of percentage of false significances over all voxels for the actual fMRI data. Opposed to GLM, SVM-GA required no spatial filtering (but tolerated it well). Moreover, SVM-GA preserved its power even with lower number of subjects.
Conclusions
SVM-GA is a powerful and robust alternative to conventional fMRI group analysis by GLM, preserving spatial resolution of fMRI data, and providing additional subject-level reports.
