Enhancing Dyslexia Detection: A Robust Machine Learning Approach

Dyslexia, recognized as a common learning disability, adversely affects reading, spelling, and word decoding capabilities, frequently resulting in considerable academic and social obstacles if not identified at an early stage. Conventional diagnostic approaches, encompassing neuroimaging and standardized assessments, prove to be effective yet are characterized by high costs and limited scalability. In response to this challenge, the present study introduces a hybrid classification model that integrates Random Forest and XGBoost methodologies to improve dyslexia detection through binary classification. The dataset employed comprises both numerical and categorical variables, such as “Gender” and “Native Language,” which were processed via one-hot encoding to facilitate the transformation of categorical information. The Synthetic Minority Oversampling Technique (SMOTE) was utilized to mitigate class imbalance, while Principal Component Analysis (PCA) was implemented to diminish dimensionality, thereby enhancing model performance. Preliminary assessments of individual classifiers indicated that Random Forest surpassed both Decision Tree and Neural Network (MLPClassifier), achieving an accuracy of 95.31%an F1-score of 0.95, and precision metrics of 0.94 for class 1 and 0.97 for class 0. Expanding upon these findings, a hybrid model that merges Random Forest and XGBoost through the application of a VotingClassifier was constructed. This hybrid framework attained an elevated accuracy of 96.23%, accompanied by consistently robust precision, recall, and F1-scores of 0.96. The results underscore the efficacy of the hybrid methodology in offering a scalable and dependable solution for the early detection of dyslexia within both clinical and educational contexts.