New Method Enables Dyslexia Detection within Minutes
The method employs AI and eye-tracking data
HSE scientists have developed a novel method for detecting dyslexia in primary school students. It relies on a combination of machine learning algorithms, technology for recording eye movements during reading, and demographic data. The new method enables more accurate and faster detection of reading disorders, even at early stages, compared to traditional diagnostic assessments. The results have been published in PLOS ONE.
Dyslexia is a condition characterised by impaired reading skills despite normal intelligence and language abilities. Dyslexia is not a disease but rather a characteristic of the brain and, to the best of our knowledge, cannot be cured, posing a challenge for the child who perceives information differently, as well as for parents and teachers.
Existing methods for detecting dyslexia are complex and demand considerable effort from the child. They involve the use of lengthy language tests, which young children are not always able to endure until completion. Researchers at the Centre for Language and Brain, the AI Research Centre, and the Faculty of Computer Science of HSE University have proposed shifting focus to other parameters, such as eye movement during reading and demographic characteristics. The scientists also applied machine learning algorithms for speeding up the process of dyslexia detection.
The study used data from more than 300 primary school-age children, all of them native Russian speakers. Based on the Standardised Assessment of Reading Skills, participants were categorised into three groups: typically developed, at risk of dyslexia, and with advanced dyslexia. Then the children were asked to read a text and while they were reading, the researchers recorded their eye movements using a device known as an eye-tracker.
The primary distinction between children with dyslexia and typically reading children lies in the duration and location of their gaze fixation on words being read. While gaze fixations last for less than a second and are barely noticeable, eye-tracking technology makes the task of recording them clear and feasible.
During the study, the first-ever dataset on eye movements in Russian-speaking children was collected and annotated.
We deliberately prepared the data for machine processing and analysis using AI, a step that will significantly advance the field of dyslexia research. It is particularly valuable that we do this using the material of the Russian language.
Anastasia Lopukhina
Co-author of the paper, Research Fellow, Royal Holloway, University of London
As a result, it was possible to train AI to determine the probability of dyslexia in a child based on information about their eye movement during reading and demographic characteristics. According to the authors of the paper, it is precisely the combination of these parameters that yields the most accurate and fastest result. The scientists tested a large number of existing machine learning algorithms, including 12 classification and eight regression models, and selected four that were the most accurate and efficient. The results obtained using these four models aligned with the outcomes of traditional diagnostics.
The top three important features for determining dyslexia were the child's gender, IQ, and age. In eye movement patterns, vertical movement appeared to be more important than horizontal movement along the lines. Children with dyslexia were found to experience a standard three-year delay in reading performance compared to their typically developing peers.
The machine learning approach holds enormous potential. It reduces the cost and streamlines the process of diagnosing dyslexia. Thanks to artificial intelligence, experts will be able to detect dyslexia within 10 to 30 minutes and dedicate the time saved to working with the child, helping them adjust to the educational system. Currently, some experts are sceptical about the development of machine methods, while others hold unrealistic expectations from artificial intelligence. Our laboratory is working to develop fast, reliable, and accurate diagnostic tools that serve the needs of professionals and children alike.
Soroosh Shalileh
Co-author of the study, Head, Laboratory of Artificial Intelligence for Cognitive Science, HSE University
Despite the advantages of using demographic data, the researchers acknowledge the necessity of reducing its volume to further enhance the method's accuracy and accessibility.
A significant volume of qualitative data greatly aids machine learning; however, it also places higher demands on the diagnostic tool and may require families to disclose additional personal information or undergo more tests. We will strive to obtain accurate results based solely on information about eye movement and reading speed.
Dmitry Ignatov
Co-author of the paper, Head, Laboratory for Models and Methods of Computational Pragmatics, Faculty of Computer Science, HSE University
Based on the new method, the Dyslector application has been developed for desktop computers with Windows and MacOS operating systems and mobile devices running on Android and iOS. The application is available for download upon request.
IQ
Olga Dragoy
Co-author of the study, Director, Centre for Language and Brain, HSE University