The Research

Our method originates from a longitudinal study known as Kronobergsprojektet. The study, which was launched almost 30 years ago, examined reading difficulties and tracked the eye movements of around 100 children — both those with difficulties and those without.

Eye-tracking during reading was conducted on children in 3rd grade, and reading difficulties were assessed until adulthood. What makes the Kronobergsprojektet study unique is: 1) children were monitored over a long period of time and 2) the integrity of the recorded eye movement data. Something that could not be done at the time but which is possible now— advanced statistical analysis — is the key to our method.

By analyzing eye movement patterns from the Kronobergsprojektet study, we were able to show that statistical models based on the data could, with a high degree of accuracy (95.6%), predict which students were experiencing difficulties after as little as 30 seconds of reading (Figure 1). The analysis also showed that the method provides a good balance between sensitivity (ability to find all of the children experiencing problems) and specificity (ability to distinguish which children were not experiencing problems). The cumulative results, which were published in PLoS One (Benfatto et al., 2016), demonstrated the potential of the method in terms of being used for screening, but also indicated that additional research was required to examine the potential further.

Figure 1. Accuracy as a function of the number of eye movement parameters when classifying recordings from the Kronobergsprojektet study (Benfatto et al., 2016).

The method was developed and evaluated in an extensive research project on dyslexia (Dyslexiprojektet). Over 3,000 children in grades 1-3 were screened over the course of two years in Järfälla Municipality and Trosa Municipality. The aim of the project was twofold: 1) to produce screening procedures that were appropriate for use in schools, and 2) to ascertain the extent to which the methodology is relevant for the entire population. Unlike the Kronobergsprojektet study, there was no long-term follow-up of the children that provided a set answer; rather, all of the children were tested using standard tests on character and word strings, rapid automatized naming of letters and reading of non-words and real words.

Approximately half of the children were also tested again a year after the initial test. The analysis of the results from the Dyslexiprojektet study shows that the method works as intended, and that it can be used as early as the spring semester of 1st grade. Accuracy is still high, with a good balance between sensitivity and specificity (Table 1); however, the level of accuracy is lower than previously because the current assignment is much more difficult.

This is in part because the distribution in the material is greater since it represents all students in grades 1-3, and in part because the set answers are not as exact or precise as those in the Kronobergsprojektet study. A major advantage of the new data set is that we can provide a continual depiction of reading ability, and the material is more representative of the situation in schools today.

Table 1. Classification of reading difficulties based on the 10th percentile in grades 1-3 (Seimyr & Benfatto, 2017).

Reading ability can be described as a continuum, and we regard reading difficulties as being in the lower segment of normal distribution. Exactly where the line should be drawn is essentially arbitrary. We have decided to use the 10th percentile, which is customary in Sweden. Since we can now give a continual depiction of reading ability, we can also say more about how all students read. The results presented at the International Dyslexia Association Conference (Seimyr & Benfatto, 2017) also demonstrate that our method has high correlation with the reference range of the entire normal distribution, and that it is useful beyond merely measuring reading speed, for example.

Figure 2. Correspondence between predicted and observed reading ability is based on our method (R2 = .78) compared to only using reading speed (R2 = .48) (Seimyr & Benfatto, 2017).