Feature selection for neuroimaging applied to word-category identification in dyslexic children

Abstract

Dyslexia is a developmental reading disorder characterized by persistent difficulty to learn how to read fluently despite normal cognitive abilities. It is a complex learning difficulty that is often hard to quantify. Traditional methods based on questionnaires are not only imprecise in quantifying dyslexia, they are also not very accurate in diagnosing it. Consequently, we aim to investigate the neural underpinnings of this reading disorder in children and teenagers, as part of a project that aims to unravel some of the neurological causes of dyslexia among children at preliteracy age. In this dissertation, we develop a study of brain activation within functional MRI scans taken when children carried out pseudo-word tasks. Our study expands recently developed machine learning-based techniques that identify which type of word the study participants were reading based solely on participant’s brain activation. Because such functional MRI data contains about 30,000 voxels, we try several feature selection techniques for removing voxels that are not very helpful for the machine learning algorithm.This procedure is widely used for maximizing the machine learning algorithm accuracy, and some of these feature selection approaches allowed us to achieve very accurate results.