Temporal Feature Selection for fMRI Analysis

Recent work in neuroimaging has shown that it is possible to classify cognitive states from functional magnetic resonance images (fMRI). Machine learning classifiers such as Gaussian Naive Bayes, Support Vector Machines, and Nearest Neighbors have all been applied successfully to this domain. Although it is a natural question to ask which classifiers work best, research has shown that the accuracy of a classifier is intimately tied to the underlying feature selection (or generation) method. Most of these feature selection methods search spatially for voxels that discriminate classes well. An empirical analysis shows, however, that voxels that discriminate well at a given time point may not discriminate well at another time point. Thus without considering this temporal component we risk passing more noise to the classifier than necessary. Choosing features temporally, focusing on regions of time when voxels discriminate well, can reduce this noise. We present empirical results that show that this method yields highly accurate classifiers with far fewer features than methods that only consider spacial information.