Using Advanced Machine Learning Models for Detection of Dyslexia Among Children By Parents: A Study from Screening to Diagnosis

Abstract

Parents of children with dyslexia have an important role in the detection and treatment of success in their children. However, standard scales in this context are not suitable for use among parents. The main aim of the current study was to find the most important indicators of dyslexia according to parents’ reports and statements. First, a list of parent reports on dyslexia was developed. Then, according to the DSM-5 criteria (by clinicians), children were divided into two categories: children with dyslexia and healthy controls. Then, four Machine Learning (ML) algorithms—Logistic Regression, Random Forest, Extreme Gradient Boosting (XGBoost), and ensemble methods—were used to extract the most relevant predictors. To predict dyslexia, recursive feature elimination chose the five most important variables from 35 parent-reported items. Logistic Regression, Random Forest, XGBoost, and ensemble models were used in R-Studio. The ensemble model was the best. The most important were “Word Guessing,” “Letter Confusion,” “Letter–Sound Association,” “Slow Reading,” and “Letter Order Reversal.” The study revealed that ML models can accurately identify dyslexia by analyzing parent-reported indicators. The five key predictors “Word Guessing,” “Letter Confusion,” “Letter–Sound Association,” “Slow Reading,” and “Letter Order Reversal” provide essential information for detecting dyslexia early.