Screening Dyslexia for English Using HCI Measures and Machine Learning
Luz Rello, Enrique Romero, Maria Rauschenberger, Abdullah Ali, Kristin Williams, Jeffrey P. Bigham, Nancy Cushen White · 2018 · International Conference on Digital Health · doi:10.1145/3194658.3194675
Summary
This paper presents a machine learning approach to screening for dyslexia in English speakers by analyzing how users interact with a linguistic computer-based game called Dytective. More than 10% of the population has dyslexia, but most are diagnosed only after failing in school, and traditional diagnostic procedures are expensive and require professional oversight. The researchers designed the game's exercises based on two foundations: empirical linguistic analysis of 833 confusion sets (groups of words commonly confused by people with dyslexia in English writing), and four cognitive skill domains related to dyslexia — Language Skills (alphabetic, phonological, syllabic, lexical, morphological, syntactic, semantic, and orthographic awareness), Working Memory (visual, auditory, and sequential), Executive Functions (activation/attention, sustained attention, simultaneous attention), and Perceptual Processes (visual and auditory discrimination and categorization). The game presents 37 levels of exercises in a Whac-A-Mole style interface, where players identify targets among carefully chosen distractors within 25-second time windows. The game was built as a cross-platform web application (HTML5, CSS, JavaScript) using dyslexia-friendly design principles: black text, mono-spaced Courier typeface, minimum 14-point font size. From each participant's gameplay, 226 features were extracted, including demographic data and six performance measures (clicks, hits, misses, score, accuracy, miss rate) per level.
Key findings
The study recruited 267 participants aged 7 to 60 from specialized centers and schools in the USA, including 52 diagnosed with dyslexia, 206 without dyslexia, and 9 at risk or suspected. Using a Gaussian Support Vector Machine (SVM) with 10-fold cross-validation, the model achieved 84.62% overall accuracy in predicting dyslexia from gameplay features. Breaking down by class: precision for dyslexia detection was 63.76% with 80.24% recall, while precision for non-dyslexia was 93.88% with 85.83% recall. This means the model correctly identifies most people with dyslexia (80% sensitivity) while maintaining high specificity for non-dyslexic readers (94% precision). The approach builds on the team's earlier Spanish-language screener (Dytective for Spanish) which achieved 83% accuracy with SVMs and later 91.97% with LSTMs on a larger dataset of 4,335 participants. The English version extends this work by incorporating a wider range of cognitive indicators beyond just linguistic features.
Relevance
This research demonstrates the potential of using HCI interaction measures as digital biomarkers for learning disabilities, making screening accessible to anyone with a computer rather than requiring expensive professional evaluation. For accessibility practitioners, the work illustrates how game-based interfaces can serve dual purposes — engaging users (especially children) while collecting clinically meaningful data about their cognitive processing. The approach has been deployed as the free online tool Dytective (dytectivetest.org), which has been used over 100,000 times, showing real-world scalability. The study is particularly relevant to digital accessibility because dyslexia significantly affects how people interact with text-heavy digital content, and early identification enables earlier intervention and accommodation. The careful game design — incorporating dyslexia-friendly typography, linguistically principled distractors, and exercises targeting specific cognitive pathways — provides a model for how HCI researchers can design assessment tools that are both scientifically rigorous and accessible to the populations they aim to serve.
Tags: dyslexia · machine learning · screening · serious games · early detection · cognitive accessibility · learning disabilities · digital biomarkers
Standards referenced: DSM-5