Particle Swarm Feature Selection for fMRI Pattern Classification

The application of pattern recognition to functional magnetic resonance imaging (fMRI) data enables exiting possibilities, including mind-reading and brain-machine interfacing. This paper presents a novel brain state identification approach, which, using an algorithm based on particle swarm optimization (PSO) in conjunction with a classifier of choice, identifies important brain voxels – thus both maximizing the classification performance and identifying physiologically relevant areas of the brain. For classifiers, we have investigated simple multiple linear regression (MLR) with thresholding and linear support vector machines (SVMs). Applying the PSO algorithm to single-subject, 2D data from a pleasant touch study, originally containing 5650 voxels, voxel subsets of mean size 64.8 and 132.6 voxels with classification accuracies of 73.1% and 77.0%, respectively for MLR and SMVs, was obtained. Similarly, on group level 3D data from a fingertapping study, with a total volume of 61078 voxels, a classification score of 83.5% was achieved on 89 voxels using the linear regression approach. For both datasets, the identified voxels agreed well with both general linear model T-maps and physiologically expected regions of activation. The PSO is thus effective in the identification of high-performing voxel subsets for fMRI volume classification, and also provides physiological information about brain processing related to the experimental conditions. Moreover, the PSO is a user-friendly algorithm, requiring little input from the user in terms of parameter specification.