Supervised-learning methods for pattern recognition in fMRI data for the identification of informative brain regions in psychological contexts

Abstract

In the last years, there has been an exponential increase in the use of multivariate analysis in neuroimaging data. This has led to a new perspective in the study of brain function, which lets identify the brain areas involved in a cognitive function and the characterization of the patterns of information associated with them. This kind of analyses are based on a classification framework that increases the sensitivity of classic univariate approaches and detects subtle changes in neural activity of different experimental conditions. This thesis focuses on three main aspects of the classification pipeline. First, an optimal estimation of the activation patterns to isolate the contribution of each experimental condition to the hemodynamic response. We have employed a Least-Squares Separate method, an approach that iteratively fits a new model for each trial in the experiment. Each model has two regressors: one for the target trial and another one for the rest. We show for the first time that this method can effectively identify the activity associated with different events even though they belong to different cognitive processes with a substantial difference in their duration. Second, a classification algorithm based on Multiple Kernel Learning (MKL). This approach combines different brain regions from an atlas to build the classification function and identifies the relevance of each region in a specific psychological context. We propose a modification of this algorithm and employ an L2-regularization to avoid the sparsity that L1-regularization entails. This approach yields simultaneously the high sensitivity of multivariate methods and the directionality of univariate approaches. Third, we employed a non-parametric approach based on permutation testing that computes more accurately the significance thresholds for each voxel in the brain. Hence, it takes into account the differences in maximum decoding accuracy that different brain regions have, enhancing sensitivity and detecting true informative that otherwise would not be marked as significant. Our results show that differences between parametric and non-parametric methods can be much larger when trying to detect subtle changes in neural activity. We have shown that information derived from just above-chance accuracies should not be underestimated if these accuracies are significant. We should expect high values of accuracy when contrasting stimuli with large perceptual differences. If these differences are minimal, the accuracy will be small. Future research should continue the development of machine learning methods especially optimized for Cognitive Neuroscience, where obtaining large accuracies is not of first interest but to provide a clear interpretation of the mathematics behind these algorithms.