Detection of developmental dyslexia with machine learning using eye movement data

Abstract

Dyslexia is a common neurocognitive learning disorder that can seriously hinder individuals’ aspirations if not detected and treated early. Instead of costly diagnostic assessment made by experts, in the near future dyslexia might be identified with ease by automated analysis of eye movements during reading provided by embedded eye tracking technology. However, the diagnostic machine learning methods need to be optimized first. Previous studies with machine learning have been quite successful in identifying dyslexic readers, however, using contrasting groups with large performance differences between diagnosed and good readers. A practical challenge is to identify also individuals with borderline skills. Here, machine learning methods were used to identify individuals with low performance of reading fluency (below 10 percentile from a normal distribution) using their eye movement recordings of reading. Random Forest was used to select most important eye movement features to be used as input to a Support Vector Machine classifier. This hybrid method was capable of reliably identifying dysfluent readers and it also provided insight into the data used. Our best model achieved accuracy of 89.7% with recall of 84.8%. Our results thus establish groundwork for automatic detection of dyslexia in a natural reading situation.