Early Prediction of Dyslexia Risk Factors in Kids Through Machine Learning Techniques

Abstract

Dyslexia is a disability that prevents people from learning to read even when they have the appropriate learning environment, education, and sociocultural environment. Dyslexia affects a person's reading skills, hinders academic success, and has long-term effects extending beyond the learning years. An early diagnosis is imperative. A series of tests are conducted to determine whether the child needs a specific set of educational techniques for learning, human experts usually evaluate these tests, and inconsistencies may also result from this human evaluation. As a result, there is a critical need for dyslexia screening that is faster, simpler, and less expensive. This study explores the feasibility of automating this screening using modern machine learning techniques. A web-game-based open-source dataset with 196 features was used. The Synthetic Minority over sampling Technique (SMOTE) was applied to balance the sample distribution. The PCA and XGBoost techniques were utilized to select dominant features. To detect dyslexic and non-dyslexic classes, various machine learning classifiers like SVM, Naive Bayes, Logistic Regression, Decision Trees, and Neural Networks were trained using 5-fold cross-validation experiments. Our Neural Network-based model achieved the highest accuracy of 95.32% with 75 features and 94.47% with nine significant features. The RBF-Support Vector Machine binary classifier achieved the highest accuracy of 96.69% with 196 features.