Representational similarity analysis with multiple models and cross-validation in magnetoencephalography

Abstract

Due to the increased availability of computational resources, more complex analysis methods taking advantage of the inherent high dimensionality of the data can be employed in functional brain imaging, allowing for development and assessment of intricate models. Models are utilized for both explanatory and predictive purposes and permits generalization from individual brain responses to the functioning principles of the brain. Representational similarity analysis (RSA) is a framework allowing evaluation of the performance of models by comparison to imaging data via the use of representational distance matrices (RDMs). This type of analysis also enables finding the linear combination of models that best explains the imaging data, something that successfully has been applied to functional magnetic resonance imaging (fMRI) data. In this thesis, RSA is applied to magnetoencephalography (MEG) data on the sensor level using a spatiotemporal searchlight approach. The method is validated through simulations based on the forward-inverse modelling framework of MEG, where complete control over the source activation is exerted. Non-negative least squares fitting of a linear combination of multiple models is carried out, with an additional option of performing leave-$k$-out cross-validation to prevent overfitting to the simulated dataset. Finally, the method is applied to real MEG data.