Professional Documents
Culture Documents
Dyslexia Deep Clustering Using Webcam-Based Eye Tracking
Dyslexia Deep Clustering Using Webcam-Based Eye Tracking
Corresponding Author:
Mohamed Ikermane
Laboratory LABSI, Faculty of Sciences, Ibn Zohr University
Agadir, Morocco
Email: [email protected]
1. INTRODUCTION
Dyslexia is a neurodevelopmental issue that causes difficulty learning to read. It decreases reading
performance and fluency and is connected with specific impairments in word immediate recognition [1].
Dyslexia influences a person's reading ability in that he may suffer to decode words, understanding vocabulary,
and process proper spelling, which has a huge negative impact on his educational outcomes [2]; therefore,
school failure is associated with dyslexia, despite the fact that it has no impact on a child's intellect [3]. In the
Organisation for Economic Co-operation and Development OECD's recent International educational survey
(PISA) [4], Morocco is ranked in the last fifth position. Students in Morocco scored lower than the OECD
average in reading performance. School failure may have many reasons like economic, social problems,
psychological and emotional problems and also physical illness and neurodevelopmental disorders reasons.
Dyslexia is regarded in Morocco as one of the most common and most damaging learning disorders
in terms of academic success. Thus, instructors must be better trained to assist students who are not progressing;
they must also be trained in diagnostic tools, remedial approaches, and accompaniment [5]. Dyslexia accounts
for 37% of Moroccan children's school failure [6]. Early dyslexia detection is determining. When diagnosed,
children with dyslexia can avoid a high rate of academic failure due to this learning disability, but identifying
it is not an easy task. There are several common ways of diagnosing Dyslexia such as WISC-V [7] in which
we solely test individual cognitive processes on which natural reading is based. The result assesses the
proportion of right replies and provides an evaluation of success on a specific reading assignment; However,
this does not represent the natural reading process [8].
Analyzing approaches based on eye-tracking throughout the reading process provides an assessment
of the reading process as it occurs naturally [9]. Many studies in this field have indicated that dyslexics'
abnormal eye movements while reading are not only due to their reading difficulties. They are, in fact, rather
independent of the reading material, and a comparison of dyslexics and normal readers reveals that eye
movement patterns and features can distinguish dyslexics from these groups of readers [10]–[12]. Thus,
screening for dyslexia using eye tracking measurements (Blink, Fixation, Saccade...) and machine learning
techniques could be extremely beneficial in identifying individuals at risk of dyslexia during the early school
years [13]–[15], which can have a significant impact on their academic achievements. This sort of screening
employs a device known as an "Eye Tracker," which is an expensive piece of equipment that only works with
specified software in some cases. As a result, utilizing eye-tracking technology to identify dyslexia may be
costly and complicated. The importance of early dyslexia diagnosis cannot be overstated; a model of dyslexia
detection via webcam-based eye tracking may be viewed as a low-cost option that tries to help overcome the
expense and limitations of employing an "Eye Tracker" while maintaining high classification accuracy. Such
a primary Dyslexia diagnosis technique can be used among students with poor reading performance to identify
individuals at high risk of dyslexia. Dyslexic readers experience longer fixations, more fixations, and shorter
saccades than non-dyslexic readers [10].
3. METHOD
3.1. Stimuli and design
This experiment was carried out through a Web-Based Application. Subjects would be given Arabic
normal text materials with various text configurations such as font-sizes, character spacing, and font family to
analyze their reading behavior. Participants sat 50 cm away from a 14-inch Lenovo ThinkPad T490 computer
with a QHD screen display of 2560x1440px resolution and a refresh rate of 60hz. A Logitech C270 HD
WEBCAM 720p/30 fps HD video with diagonal 55° field of view and auto light correction eye tracking their
gaze points. The web-based app was written in HTML5/CSS3 and ran on an Apache Server 2.4.51. Owing to
the clear documentation and simplicity of integration into a new application, the eye-tracking processes were
modified from the eye-tracking WebGazer algorithm [23]. We primarily adjust it to our requirements by
employing a clear webcam-based eye-tracking approach. The experimenter must fill out some personal
information about the subject before beginning the experiment as illustrated in Figure 1, including the subject’s
name, age and sex, school grade, and reading level.
Figure 1. Home page of the web-application where we insert the participant’s personal information
language. The content is made up of seven paragraphs and a total of 155 words, with no numerical expressions,
acronyms, or foreign words Table 1.
To avoid sequence effects, the text presentation was also controlled. The content and fonts were
counterbalanced. As a result, the text paragraphs were divided into two pages, the first of which contained four
paragraphs (84 Words) with the most common fonts used on screen and in printed texts [26] Arial, Arial Italic,
Times, and Times Italic for each paragraph. All of the texts were center-justified, with a font size of 18 points
and with black text color on a white background.
The second page of the paper utilized Arabic dyslexic-friendly typefaces to improve readability for
dyslexic readers. Specifically, three different Arabic typefaces were used: Arabic transparent, simplified
Arabic, and simplified Arabic fixed fonts with a size of 20 points. These typefaces were formatted using a
combination of bold and black color, which is recommended for Arabic web content. Overall, the use of
dyslexic-friendly typefaces and formatting can have a significant impact on the readability and comprehension
of text for individuals with dyslexia [27].
To evaluate the performance of our Autoencoder, we divided our dataset into a training set (with a
ratio of 0.75) and a test set (with a ratio of 0.25). From Figure 10, which displays the loss of our Autoencoder
over the training and test data, we can infer that the ANN performed well in coding and decoding our dataset
inputs. Unsupervised learning involves the challenge of determining the optimal number of clusters, and to
tackle this issue, we utilized the elbow method and silhouette analysis techniques in K-means clustering to
identify the most suitable number of clusters that best fits our eye-tracking dataset.
Figure 10. The autoencoder model loss over training and test data
common graphical representation of the elbow method, showing how the SSE values change with the number
of clusters and highlighting the elbow point as a candidate for the optimal number of clusters.
The term "silhouette" refers to a way of interpreting and validating consistency inside data clusters, it
can also assist in determining the appropriate k value. used to investigate the separation distance between the
clusters that resulted. The silhouette plot shows how near each point in one cluster is to points in neighboring
clusters, the range of this metric is [-1, 1]. Results showed that 3 is the best K value of our dataset with a
silhouette score of 0,65. Analyzing the elbow curve and the silhouette scores we can determine that K=3 is the
number of clusters that fit best our dataset. When evaluating our three clusters, the first one is represented by
shorter saccades, longer duration of fixation, and more fixations compared to the other two clusters, also this
group of participants benefited the most from using Arabic dyslexic-friendly typefaces when comparing the
eye-tracking reading outcomes for the two texts. The mean duration of fixation has decreased for all members
of this cluster, ranging between 24% and 60%, and saccades have also increased, with an average of 63% to
81%. This cluster was identified as “High Risk of Dyslexia”. The second cluster was less affected by Dyslexia
traits with fewer fixations, long saccades, and shorter fixations. we categorized these cluster members as
"Typical Readers". We labeled the third cluster as "Risk of Dyslexia", participants also showed longer fixation
durations but fewer fixations and saccades than the first group. Figure 12 illustrates our eye-tracking dataset
clustering using K-means and Deep clustering technics.
5. CONCLUSION
Dyslexic people's eye motions differ from those of normal readers. People with dyslexia tend to have
longer fixations, longer reading periods, and more fixations than typical readers. In this research, we
demonstrated how these traits, in conjunction with machine learning and webcam-based eye-tracking may be
used to identify dyslexia. We got promising results by demonstrating that the use of Arabic Dyslexia friendly
typefaces such as Arabic transparent, simplified Arabic, and simplified Arabic fixed fonts with significant font
size could improve reading performance, and that poor readers benefited the most from the use of these
typefaces by reducing fixation number and duration and expanding saccades Furthermore, the results
confirmed that webcam-based eye-tracking technics can be used as an assessing tool for investigating the
student's reading performance and the potential risk of dyslexia. To summarize, we utilized a JavaScript-based
eye-tracking framework and consumer-grade webcams to record participants' eye movements and gaze
positions during reading Arabic text using a web Application and Deep Clustering technics to extract the best
features and reduce the dimensionality of our eye-tracking data collection. then we applied the k-means
clustering method to estimate Dyslexia prevalence among students.
ACKNOWLEDGEMENTS
The authors thank Primary school teachers and directors for their cooperation.
REFERENCES
[1] I. C. J. Adubasim and J. T. Nganji, “Dyslexia-A Learning Difference,” Autism Opne Access, vol. 7, no. 1, pp. 1–4, 2017, [Online].
Available: https://www.omicsgroup.org/journals/the-rapidly-changing-landscape-of-multiple-sclerosisimmunomodulatory-
therapy-a-retrospective-chart-review-in-the-unitedarab-emirat-2376-0389-1000203.php?aid=90534.
[2] M. J. Snowling, “1. Does dyslexia exist?,” Dyslexia: A Very Short Introduction, pp. 1–11, 2019,
doi: 10.1093/actrade/9780198818304.003.0001.
[3] S. Gustafson and S. Samuelsson, “Intelligence and dyslexia: Implications for diagnosis and intervention,” Scandinavian Journal of
Psychology, vol. 40, no. 2, pp. 127–134, 1999, doi: 10.1111/1467-9450.00109.
[4] OECD, 21st-Century Readers: Developing literacy skills in a digital world. OECD, 2021.
[5] S. Ihbour, “Dyslexia in Morocco: An overview of recent research,” IOSR Journal of Research & Method in Education
(IOSR-JRME), vol. 9, no. 1, pp. 24–26, 2019, doi: 10.9790/7388-0901012426.
[6] M. Abdellah, “Invisible disabilities between education and health.”
[7] D. Study Citation Wechsler, “Technical and interpretive manual supplement,” Pearson, p. 84, 2014.
[8] M. N. Benfatto, G. Ö. Seimyr, J. Ygge, T. Pansell, A. Rydberg, and C. Jacobson, “Screening for dyslexia using eye tracking during
reading,” PLoS ONE, vol. 11, no. 12, 2016, doi: 10.1371/journal.pone.0165508.
[9] J. Everatt, M. F. Bradshaw, and P. B. Hibbard, “Individual differences in reading and eye movement control,” Eye Guidance in
Reading and Scene Perception, pp. 223–242, 1998, doi: 10.1016/b978-008043361-5/50011-0.
[10] G. F. Eden, J. F. Stein, H. M. Wood, and F. B. Wood, “Differences in eye movements and reading problems in dyslexic and normal
children,” Vision Research, vol. 34, no. 10, pp. 1345–1358, 1994, doi: 10.1016/0042-6989(94)90209-7.
[11] J. Hyönä and R. K. Olson, “Eye fixation patterns among dyslexic and normal readers: Effects of word length and word frequency,”
Journal of Experimental Psychology: Learning, Memory, and Cognition, vol. 21, no. 6, pp. 1430–1440, 1995, doi: 10.1037/0278-
7393.21.6.1430.
[12] R. K. Olson, “Eye movements in dyslexic and normal readers,” Journal of Experimental Psychology: Human Perception and
Performance, vol. 9, no. 5, pp. 816–825, 1983.
[13] L. Rello and M. Ballesteros, “Detecting readers with dyslexia using machine learning with eye tracking measures,” W4A 2015 -
12th Web for All Conference, 2015, doi: 10.1145/2745555.2746644.
[14] A. Jothi Prabha and R. Bhargavi, “Prediction of dyslexia from eye movements using machine learning,” IETE Journal of Research,
vol. 68, no. 2, pp. 814–823, 2022, doi: 10.1080/03772063.2019.1622461.
[15] T. Asvestopoulou et al., “DysLexML: Screening tool for dyslexia using machine learning,” 2019, [Online]. Available:
http://arxiv.org/abs/1903.06274.
[16] F. AlRowais, M. Wald, and G. B. Wills, “An Arabic framework for dyslexia training tools,” 1st International Conference on
Technology for Helping People with Special Needs (ICTHP-2013) (19/02/13 - 20/02/13), pp. 63–68, 2013, [Online]. Available:
http://eprints.soton.ac.uk/id/eprint/353261.
[17] N. Friedmann and M. Haddad-Hanna, “Letter position dyslexia in Arabic: From form to position,” Behavioural Neurology, vol. 25,
no. 3, pp. 193–203, 2012, doi: 10.1155/2012/296974.
[18] S. A. Rabia and A. Salfeety, “Reading in Arabic orthography: The influence of short vowels on reading accuracy and
comprehension of poor and normal Arabic readers,” Journal of Advances in Linguistics, vol. 5, no. 2, pp. 723–735, 2015,
doi: 10.24297/jal.v5i2.2848.
[19] S. Abu-Rabia, “Reading Arabic texts: Effects of text type, reader type and vowelization,” Reading and Writing, vol. 10, no. 2,
pp. 105–119, 1998, doi: 10.1023/A:1007906222227.
[20] A. Al-Edaily, A. Al-Wabil, and Y. Al-Ohali, “Dyslexia explorer: A screening system for learning difficulties in the arabic language
using eye tracking,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture
Notes in Bioinformatics), vol. 7946 LNCS, pp. 831–834, 2013, doi: 10.1007/978-3-642-39062-3_63.
[21] A. Al-Edaily, A. Al-Wabil, and Y. Al-Ohali, “Interactive screening for learning difficulties: Analyzing visual patterns of
reading arabic scripts with eye tracking,” Communications in Computer and Information Science, vol. 374, no. PART II, pp. 3–7,
2013, doi: 10.1007/978-3-642-39476-8_1.
[22] T. M. Al-ghizzi, “Dyslexia in the Arab world,” International Journal of Language and Literature, vol. 3, no. 2, 2015,
doi: 10.15640/ijll.v3n2a7.
[23] A. Papoutsaki, “Scalable webcam eye tracking by learning from user interactions,” Conference on Human Factors in Computing
Systems - Proceedings, vol. 18, pp. 219–222, 2015, doi: 10.1145/2702613.2702627.
[24] G. Q. Salahia, “Constructing a test of reading ability in Arabic,” University of Algiers, 1998.
[25] P. Lefavrais, “Description, définition et mesure de la dyslexie. Utilisation du test ‘l’Alouette,’” Revue de Psychologie Appliquée,
vol. 15, pp. 33–44, 1965.
[26] L. Rello and R. Baeza-Yates, “Good fonts for dyslexia,” Proceedings of the 15th International ACM SIGACCESS Conference on
Computers and Accessibility, ASSETS 2013, 2013, doi: 10.1145/2513383.2513447.
[27] F. Benmarrakchi and J. El Kafi, “Investigating reading experience of dyslexic children through dyslexia-friendly online learning
environment,” International Journal of Information and Communication Technology Education, vol. 17, no. 1, pp. 105–119, 2021,
doi: 10.4018/IJICTE.2021010107.
[28] M. Caron, P. Bojanowski, A. Joulin, and M. Douze, “Deep clustering for unsupervised learning of visual features,” Jul. 2018,
doi: 10.1007/978-3-030-01264-99.
BIOGRAPHIES OF AUTHORS