Artificial intelligence: A data analysis system to subtype dyslexic children using electroencephalograms
Abstract
Three to six percent of school aged children in the United States suffer from some form of dyslexia. It has been reported, there are over 1400 ways to identify learning disability. In this study, neural networks were used to develop a reliable method to identify dyslexia in children. Using the Boder typology of dyslexia, fifteen children: five normals, six dysphonetic dyslexics, and three dysorthographic dyslexics were chosen for this study. Two-second segments of electroencephalogams (EEG) were recorded using the 10-20 International setup system during reading and resting states. The Multitaper Power spectral Density of EEG was calculated. The Mann-Whitney tests were used on spectral parameters to obtain variables that are able to differentiate between the three groups. The regions of difference between dyslexic subtypes and normals were in areas involved in speech functions. Based on the Mann-Whitney, frequency bands from the EEG were used as inputs to the neural network model. The neural network was able to identify all subjects with an average of 93.1% of the segments in the reading state. The neural network could not discriminate between the groups in the resting state. The neural network was compared to discriminant function analysis. The neural network classifications outperformed discriminant function analysis in segment classification (74.5%) and subject classification. The neural network was able to classify all unremediated subjects correctly while discriminant function analysis misclassified two subjects. Compared to other methods employed currently reported in literature, the neural network had 100% subject classification. Furthermore, the neural network was able to diagnose severity among dyslexic subjects.
This paper has been withdrawn.