Learning disorder detection based on rest EEG signals using time-frequency features of the discrete orthonormal...

Learning disorders are neurodevelopmental conditions that impair a person's processing abilities, such as reading, writing, and arithmetic, called dyslexia, dysgraphia, and dyscalculia, respectively. This study aims to develop an automated framework for detecting learning disorders using resting-state electroencephalogram (EEG) signals. EEG data were collected from 149 Iranian students aged 7-11 years, including healthy controls and subjects diagnosed with different learning disorders, under both eye-open and eye-close conditions. The models are trained and tested with the seen data, and their performance in the real world is evaluated with the unseen data. After preprocessing operations, time-frequency features were extracted using the discrete orthonormal Stockwell transform (DOST) and its discrete cosine variant (DCT-DOST). Then, feature selection was performed using the Wilcoxon rank-sum and Kruskal-Wallis tests, and classification was carried out with support vector machine, decision tree, adaptive boosting, and random forest (RF) classifiers, utilizing several dataset balancing methods. For the 2-class classification (control, learning disorder), the best performance was achieved using the RF classifier with synthetic minority oversampling technique (SMOTE) balancing, reaching accuracies of 94.77 ± 0.44% (seen data) and 82.17% (unseen data) for eye-open, and 93.72 ± 0.46% (seen data) and 87.27% (unseen data) for eye-close signals. For the 4-class classification (control, dyscalculia, dyslexia/dysgraphia, mixed), the same configuration achieved accuracies of 98.28 ± 0.22% (seen data) and 68.25% (unseen data) for eye-open, and 97.75 ± 0.17% (seen data) and 62.33% (unseen data) for eye-close signals. In conclusion, the proposed DOST-based framework, combined with the RF classifier and SMOTE, provides a reliable approach for the early detection of learning disorders.