Prediction of Dyslexia and its Stages Based on Eye Gaze Points

Abstract

Dyslexia is a neurological disorder characterized by difficulty in reading. Dyslexics newlinehave difficulty in phonological processing and word decoding, hence they have less newlinecapability to relate letters, form words, and also exhibit poor reading comprehension. newlineDyslexia is not a visual impairment disorder but many dyslexics have an impaired magnocellular newlinesystem which causes poor eye control leading to abnormal eye movement newlinepatterns. The literature says that disruptive eye movements could be potential indicators newlineof dyslexia. Eye-trackers can be used to record eye gaze points while reading. Early newlineprediction of dyslexia in children can help them take intervention programmes at an earlier newlinestage helping them to excel in academics. This research work focuses on the early newlineprediction of dyslexia in children from their eye gaze points while reading. Eye gaze newlinepoints are explored, analyzed, and a set of eye movement features based on fixations and newlinesaccades are proposed, that contribute more towards the prediction of dyslexia, while newlineusing machine learning models. A combinational kernel is proposed for developing newlinean improvised machine learning model, to screen dyslexics from non-dyslexics using newlinethe proposed features. This research work also analyzes the dyslexic data using the proposed newlinefeatures to identify the stages of dyslexia. Eye movement events such as fixations newlineand saccades are detected using simple statistical measures, dispersion-threshold, and newlinevelocity-threshold algorithms. A statistical t-test is conducted to know the significance newlinelevel of features. This work proposes a combinational kernel for Support Vector Machine newline(SVM) to develop an efficient prediction model for dyslexia. The proposed model newlineis validated on 185 children of 9-10 years old. Validation is done using 10-fold crossvalidation newlineand average accuracy was considered. Train and Test data are split in the newlineratio of 80% and 20% respectively. A high-level accuracy of 97.5 % was achieved using newlinethe proposed combinational kernel for SVM. The proposed model gave an impro