Screening Risk of Dyslexia Through a Web-Game Using Language-Independent Content and Machine Learning
Maria Rauschenberger, Ricardo Baeza-Yates, Luz Rello · 2020 · Proceedings of the 17th International Web for All Conference (W4A) · doi:10.1145/3371300.3383342
Summary
This paper presents MusVis, a web-based game designed to screen for dyslexia risk using language-independent content and machine learning, enabling potential early detection even in pre-readers who have not yet developed literacy skills. Dyslexia affects 5-15% of the world population and is typically diagnosed only after children fail in school, despite being unrelated to general intelligence. Current screening approaches require expensive professional assessment or language-specific reading tests that cannot be administered to young children who have not yet learned to read. The authors designed MusVis with two complementary components: an auditory part inspired by the card-matching game Memory, where children find pairs of matching sounds by clicking on cards, and a visual part using a Whac-A-Mole style interaction where children identify target visual cues (symbols, letters, rectangles, faces) among distractors. Crucially, the game content uses no linguistic material — the auditory cues are generated sine tones varying in frequency, rhythm, length, and rise time, while the visual cues use abstract shapes featuring horizontal and vertical symmetries known to be challenging for people with dyslexia. The game was built as a web application using JavaScript, jQuery, CSS, and HTML5, with a PHP/MySQL backend, and takes less than 10 minutes to play. Game mechanics include rewards (points), instant feedback, challenges (time limits), and a narrative story to maintain engagement. The study used a human-centered design framework, with game instructions translated into each participant's native language while game content remained identical across languages.
Key findings
A user study with 313 children aged 7-12 (116 diagnosed with dyslexia, 197 controls) across German and Spanish speakers yielded promising classification results. Using Random Forest classifiers for German data, the system achieved 74% accuracy and an F1-score of 0.75. For Spanish data, Extra Trees classifiers achieved 69% accuracy and an F1-score of 0.75. The best models used relatively few features — just 5 for German and 20 for Spanish — suggesting that a small number of behavioural indicators captured through gameplay are sufficient for prediction. Key distinguishing variables included total number of clicks, time to first click, efficiency, and click intervals in the visual game, and 4th click interval, duration, and average click time in the auditory game. Interestingly, children with dyslexia did not make more mistakes than the control group; instead, they showed different temporal patterns — taking more time to process information, consistent with known difficulties in auditory and visual perception. The auditory and visual features were equally represented in the most informative feature rankings, supporting the theory that dyslexia involves both auditory and visual perception differences. However, combining German and Spanish data into a single model decreased accuracy to 61%, indicating that while the game content is language-independent, the predictive models benefit from being language-specific, possibly due to cultural differences such as musical training and the transparency of orthographic systems.
Relevance
This research addresses a critical gap in early identification of learning disabilities. Children with dyslexia typically need about two years to compensate for their reading and spelling difficulties, and late diagnosis leads to school failure, anxiety, and decreased self-esteem. A web-based screening tool that requires no reading ability could enable identification before formal literacy instruction begins, facilitating early intervention when it is most effective. For accessibility practitioners, this work demonstrates how gamification and human-centered design can make screening tools engaging rather than clinical, reducing barriers to participation. The language-independent approach is particularly significant for multilingual contexts and underserved populations where professional diagnostic resources are scarce. The tool's web-based delivery means it can be deployed at scale through schools without specialized hardware. However, the moderate accuracy levels (69-74%) position this as a screening tool that flags risk rather than a diagnostic instrument — children identified as at-risk would still need professional assessment, but the tool could dramatically increase the number of children who receive that referral.
Tags: dyslexia · machine learning · screening · serious games · gamification · auditory perception · visual perception · early intervention · language-independent
Standards referenced: DSM-5