Predicting Dyslexia and Vocabulary Age
Summary
Methods from the field of machine learning can be applied to predict disorders such as dyslexia at an individual level. Early vocabulary is a plausible antecedent to later language difficulties. Previous research has been mostly correlational and not predictive. In this study, productive and receptive vocabulary sizes measured in children (n=147) between 17 and 35 months were used to predict their later dyslexia status. Linear support vector machines were trained to separate dyslexic subjects from nondyslexic subjects. Additionally, support vector regression was used to predict age from vocabulary to see if there is a significant difference in age prediction between dyslexic and nondyslexic children. Dyslexia could not be reliably predicted using age-specific classification models: the maximum balanced accuracy was 58\% for the group at 23 months old. The vocabulary age models did have a good fit (for the best model, R squared=0.686) and performed well on unseen data. Moreover, the age models predicted dyslexic subjects to be up to two months younger than their nondyslexic peers. This difference was however not enough to predict eventual dyslexia status. In conclusion, infant vocabulary was a weak predictor of dyslexia. Using data from multiple points in time might increase predictive performance, as vocabulary trajectory is nonlinear and differs in children with dyslexia. Similarly, the "vocabulary age gap" could be examined further, since vocabulary age models predicted dyslexics to be younger even without prior knowledge of dyslexia.