Abstract
Machine learning can be used to find meaningful patterns characterizing individual differences. Deploying a machine learning classifier fed by local features derived from graph analysis of electroencephalographic (EEG) data, we aimed at designing a neurobiologically-based classifier to differentiate two groups of children, one group with and the other group without dyslexia, in a robust way. We used EEG resting-state data of 29 dyslexics and 15 typical readers in grade 3, and calculated weighted connectivity matrices for multiple frequency bands using the phase lag index (PLI). From the connectivity matrices, we derived weighted connectivity graphs. A number of local network measures were computed from those graphs, and 37 False Discovery Rate (FDR) corrected features were selected as input to a Support Vector Machine (SVM) and a common K Nearest Neighbors (KNN) classifier. Cross validation was employed to assess the machine-learning performance and random shuffling to assure the performance appropriateness of the classifier and avoid features overfitting. The best performance was for the SVM using a polynomial kernel. Children were classified with 95% accuracy based on local network features from different frequency bands. The automatic classification techniques applied to EEG graph measures showed to be both robust and reliable in distinguishing between typical and dyslexic readers.
