Automatic classification of cognitive states

Abstract

Functional Magnetic Resonance Imaging has established itself as the most powerful technique available today to measure brain activity induced by a perceptual or cognitive state. The inverse problem is considered in this study; given the measured brain activity, our goal is to predict the perceptual state. Machine Learning algorithms were used to address this problem in this work. Multi-subject fMRI data analysis poses a great challenge for the machine learning paradigm, by its characteristics: the low Signal to Noise Ratio (SNR), high dimensionality, small number of examples and inter-subject variability. To address this problem, several methods of classification and feature selection were tested. The main criterion of feature selection was mutual information in a univariate method, but a multivariate feature selection was also proposed. Both a single classifier and an ensemble of classifiers were tested. The ensemble of classifiers approach consisted on training an optimized classifier for each class and then the combination was made. The data analysed was obtained from three multi-subject experiments of visual stimulation with 4 classes of stimuli, at different magnetic field strengths. The ensemble of classifiers performs best for most data sets and methods of feature selection. In summary, the results suggest that a combination of classifiers can perform better than a single classifier, particularly when decoding stimuli associated with specific brain areas.