Abstract
The ability to read is one of the most important skills that humans learn. There are vast individual differences in reading ability and this variation in reading ability can relate to neural differences, especially in white matter tracts (e.g. typical vs. dyslexic readers). Predicting reading disabilities is useful for early intervention, yet making reliable predictions is difficult. Here, we attempt to use diffusion weighted imaging (DWI) connectivity to improve the strength and accuracy of predictions at the individual subject level. We use machine learning algorithms, specifically support vector machine regressions (SVRs), to perform multivariate predictions of reading ability. Using DWI data, we ran probabilistic tractography using 86 pre-defined anatomical parcels as seeds and targets. We extracted the structural connectivity between seeds and targets in 220 subjects aged 14–16 years, and trained the SVR algorithms using a randomly selected subset of subjects (training dataset, N=120). These trained models were then tested on the remainder of the subjects (test dataset, N=100) in order to make predictions. Results showed that actual reading scores significantly correlated with predicted reading scores from the independent test dataset. We also tried the same approach but using group-average functional parcels instead of anatomical ones as seeds and targets. This model was also accurate in its predictions of reading ability. Both models outperformed control models and suggest that the connectivity of visual regions (e.g. VWFA) as well as possible language and attentional regions are predictive of reading ability. Our results demonstrate the feasibility of using DWI connectivity to predict individual reading ability. Due to the robust nature of these models, future applications may aid in the early identification of reading disabilities such as dyslexia.