Machine Learning Framework for Inferring Cognitive State from Magnetoencephalographic (MEG) Signals

We develop a robust linear classification framework for inferring mental states from electrophysiological (MEG and EEG) signals. The framework is centered around the concept of temporal evolution of regularized Fisher Linear Discriminant classifier constructed from the instantaneous signal value. The value of the regularization parameter is selected to minimize the classifier error estimated by cross-validation. In addition, we build upon the proposed framework to develop a feature selection technique. We demonstrate the framework and the feature selection technique on MEG data recorded from 10 subjects in a simple visual classification experiment. We show that using a very simple adaptive feature selection strategy yields considerable improvement of classifier accuracy over the strategy that uses fixed number of features.