Machine Learning, Web Game Help Identify Child's Risk of Dyslexia

Maria Rauschenberger, PhD

A recent analysis proved machine learning and a web-based game could help identify a child's risk of dyslexia.

Maria Rauschenberger, PhD, and a team of investigators developed the web game, MusVis, to measure the differences in the interactions of children with and without dyslexia while identifying both visual and musical elements. The team aimed to detect the condition through interactions that did not require a knowledge of language.

The findings showed a .74 accuracy for German and .69 for Spanish language, and highlighted the method could be used to detect possible learning disorders in children even before they develop language skills, which could result in early intervention.

“To our knowledge, this is the first time the risk of dyslexia has been analyzed by means of a web game based on language-independent contents and using machine learning,” Rauschenberger, from Pompeu Fabra University in Spain, said in a statement.

The team of investigators designed MusVis based on the assumption non-linguistic content such as rhythm and frequency could represent difficulties children with dyslexia have with writing and reading. They felt dyslexia could be measured through total number of clicks or play duration.

The language-independent game content was created with an auditory and visual part, each with 4 stages. Each stage had 2 rounds, which brought the total to 16 rounds for the whole game. In each stage there was a round with 4 cards and then 6 cards. The overall game took about 10 minutes to play.

The auditory part of the game was inspired by a game called “Memory,” in which pairs of identical cards were face down and had to be turned over. The team assigned each stage to 1 acoustic parameter like frequency, rhythm, rise time, and length. Each stage had 3 auditory cues.

The visual game used 'Whac-A-Mole' interaction with independent visual cues. The visual cues were designed with the potential of making more cues with similar features to represent horizontal and vertical symmetries known to be difficult for a person with dyslexia.

Participants were shown the target visual cues for 3 seconds and were asked to remember them. Then, the participants were presented with a setting whether the target cue and distractors were displayed. The participant needed to click on the target visual cue as often as possible in 15 seconds. The arrangement of the cues changed after every click.

There were 4 stages in the visual part. Each stage was assigned to 1 visual type (symbol, z, rectangle, face) and 4 visual cues for each stage. In each stage, there were 2 rounds with first, a four-squared and then a nine-squared design. The target and distractors were displayed to fit the design.

Overall, 313 participants were taken for the predictive models. A participant was removed if they reported in a background questionnaire having suspicions of dyslexia without a diagnosis. The rest of the participants were classified as diagnosed or not show any signs of dyslexia (control group).

The investigators separated their data in 3 sets: 1 for the Spanish participants (153 individuals), 1 for the German participants (149 individuals), and 1 for all participants (313 participants)—including 11 participants who spoke English.

Using extra trees, random forest, and gradient boost machine-learning models, the investigators processed the game data. The best accuracy reached was 74% for the German case using random forest and the best accuracy for Spanish was 69% using the extra trees model.

“Since children with dyslexia need about 2 years to offset their difficulties, our method, as it is language-independent, could help reduce school failure, delayed treatment and, most importantly, reduce the suffering of children and parents," Rauschenberg concluded.

Future research should include more children at an early age to expand the training data for the predictive models and improve the results.

The study, “Screening Risk of Dyslexia through a Web-Game using Language-Independent Content and Machine Learning,” was published online in W4A Proceedings.