Classifying cognitive states from fMRI data using neural networks

Since the discovery of functional magnetic resonance imaging (fMRI) studies have proved that this technique is one of the best for collecting vast quantities of data about activity of the human brain. Our aim is to use this information in order to predict the cognitive status of the subject given its fMRI activity. We present a new approach for creating single-subject classifiers using bagging from a pool of feed-forward backpropagation networks. Our experiments indicate that as the number of selected features (voxels) increases, the accuracy of the system increases too. Nevertheless, when the number of voxels exceeds 120, the accuracy of the system rapidly increases from 45% to 70%. Eventually it reaches a (near) saturation point after which the increase in the accuracy is very slow.