Motor Dysgraphia and Dyslexia
For a more technical explanation, click here
Learning disorders are widely understood as being neuro-biological in origin: as such, they remain a long-life condition. Motor Dysgraphia and Dyslexia are disorders whose complete diagnosis has to be made by a practitioner qualified to administer the range of standardized tools required for an accurate diagnosis. Nevertheless, these disorders are also characterized by specific difficulties. The book from the American Psychiatric Association also known as DSM-5 is an authoritative volume providing such characteristics:
- Motor Dysgraphia appears under the umbrella terms of Developmental Coordination Disorder (previously known as Dyspraxia), and affects fine motor skills, including handwriting skills.
- Dyslexia is characterized by problems with accurate or fluent word recognition, poor decoding, and poor spelling abilities.(DSM-5 page 105)
The above characteristics can be accurately measured by a computer. Such accurate measurements have led the Artificial Intelligence community to consider that, given enough data, accurate screenings can be performed by
- capturing appropriate data from an individual, then
- applying specific algorithms derived from mathematics and computer science which are not prone to the variability of human judgement.
Let us consider the scientific references for the interested reader:
Motor Dysgraphia scientific references
For motor dysgraphia, we can cite the research papers of Mekyska et al.(2013): Identification and Rating of Developmental Dysgraphia by Handwriting Analysis, and more recently Asselborn et al. (2018): Automated human-level diagnosis of dysgraphia using a consumer tablet, Zolna et al. (2019) The Dynamics of Handwriting Improves the Automated Diagnosis of Dysgraphia. The main philosophy is to capture handwritten data (which could be letters or words or even sentences), from both dysgraphic and non-dysgraphic children, then to apply classical machine learning classification algorithms.
A complete software has also been developed by Guinet and Kandel in 2010: Ductus: A software package for the study of handwriting production. More recently, the work of Asselborn et al. has been the scientific basis for a Swiss company named dynamico.ch, also supported by the Chili lab (University of Lausanne).
Dyslexia scientific references
Scientists have also investigated the use of AI techniques for dyslexia screening purposes. For instance in Features and machine learning for correlating and classifying between brain areas and dyslexia (2018), A. Frid and L. Manevitz getting a child to read 96 real words and 96 nonsense words, then analyzing their brain activity via electro-encephalograms. Then they apply a classical machine learning algorithm to get a very accurate classifier. With the same perspective, Machine learning and dyslexia: Classification of individual structural neuro-imaging scans of students with and without dyslexia. from P. Tamboer et al. (2016) investigate brain scans to detect dyslexia.
From another angle, Detecting readers with dyslexia using machine learning with eye tracking measures. from L. Rello and M. Ballesteros (2015), then M. Benfatto et al. Screening for Dyslexia Using Eye Tracking during Reading (2016) have developed an eye-tracking process paving the way to an effective process currently distributed via a successful company Lexplore.
The list is quite long: all these works provide strong support for the use of Machine Learning as an effective tool to help screen dyslexia. At Dystech, we are building on these approaches having in mind to build a screening as less invasive as possible. A complete description of our approach is described in our paper Dyslexia and Dysgraphia prediction: A new machine learning approach (2020).
What we measure for Motor Dysgraphia
During the screening, we ask for a picture of handwritten text, normally 4 to 5 lines are enough. It is recommended to follow our instructions for taking the picture as this will simplify our processing and make our measurement more accurate. Like the works mentioned above, we do not check the spelling of the text as it does not seem relevant for screening motor dysgraphia. From the text, we extract 8 features:
- Slant: The slant corresponds to the direction of the handwriting. We normalize this by converting the values from 0 (left slant) to 1 (right slant).
- Pressure: The estimated pressure of the handwriting from 0 (low) to 1 (high).
- Amplitude: This is the average gap size between x-height and ascending/descending letters, from 0 (low) to 1 (high).
- Letter Spacing: This estimates the average spacing between letters in a word. Typically, a cursive writing style will lead to 0, from 0 (small spacing 0 to 1 (large spacing).
- Word spacing: This estimates the average spacing between words in a sentence, from 0 (small spacing) to 1 (large spacing).
- Slant Regularity: from 0 (not regular) to 1 (highly regular).
- Size Regularity: from 0 (not regular) to 1 (highly regular). Measure if the same letters vary in size within the text.
- Horizontal Regularity: from 0 (the text doesn’t follow a horizontal line) to 1 (the text follows a horizontal line).
Here is an example for slant feature:
We have manually extracted these measurements from every picture we have received (more or less 1500), coming from standard writers and dysgraphic writers. Then classical processing and machine learning methods have been applied to finally provide an algorithm able to:
- automatically extract the measurements from a new picture,
- provide a likelihood of motor dysgraphia for the individual providing this new picture.
What we measure for Dyslexia
During the screening, we ask to read a list of 32 words and record the readings. 16 words are real English words and 16 are nonsense words, also generated by Artificial Intelligence techniques based on the age of the individual being screened. The complete list of words is generated in accordance with the age of the individual. The audio recordings are processed on our server via specific software and we extract 6 metrics which are important for the likelihood estimate:
- 3 metrics are related to what is known as Reading Reaction Time (RRT). For a given word, the RRT is the interval between the initial display of the word and the start of the reading. We get an average RRT on all words, on real words and on nonsense words.
- 3 other metrics are related to Reading Time (RT) which is an indication of the time it takes to read a word. We get an average RT on all words, on real words and on nonsense words.
As can be seen on the figure below (based on our data), these reading times highlight a clear discrepancy between standard readers and dyslexic readers (The unit of time we use is millisecond (ms)):
We have trained our machine learning algorithm on data coming both from standard readers and dyslexic readers (that we have obtained from educational & clinical partners). Analyzing the 32 audio records coming from a new individual, our final classifier is able to provide an accurate likelihood of dyslexia.
We will be more than happy to provide more references and explanations to interested people. Just drop us an email at firstname.lastname@example.org.