Diagnosing Dyslexia in Early School-Aged Children Using the LSTM Network and Eye Tracking Technology

Abstract

Dyslexia, often referred to as a specific reading disability, affects many students around the world. It is a neurological disorder that affects the ability to recognise words, and it causes difficulties in writing and reading comprehension. Previous computer-based methods for the automatic detection of dyslexia in children have had low efficiency due to the complexity of the test administration process and the low measurement reliability of the attention measures used. This paper proposes the use of a student’s mobile device to record the spatio-temporal trajectory of attention, which is then analysed by deep neural network long short-term memory (LSTM). The study involved 145 participants (66 girls and 79 boys), all of whom were children aged 9 years. The input signal for the neural network consisted of recorded observation sessions, which were packets containing the child’s spatio-temporal attention trajectories generated during task performance. The training set was developed using stimuli from Benton tests and an expert opinion from a specialist in early childhood psychology. The coefficients of determination of $R2\sim 0.992$ were obtained for the proposed model, giving an accuracy of 97.7% for the test set. The ease of implementation of this approach in school settings and its non-stressful nature make it suitable for use with children of different ages and developmental stages, including those who have not yet learned to read. This enables early intervention, which is essential for effective educational and emotional support for children with dyslexia.

1. Introduction

In the information age we live in, technology and data have become key elements of people’s daily lives. Two areas that can benefit from data collection are education and early educational diagnostics, especially in the context of the often-occurring problem of dyslexia. Dyslexia, often referred to as a specific reading disorder, affects many students worldwide. It is a neurological disorder that impacts the ability to recognize words, causing difficulties in writing and understanding text. The effects of dyslexia are not limited to the classroom but can also lead to problems with self-esteem, frustration, and social isolation. As experts in dyslexia diagnosis point out [1], it is a neurological disorder affecting the ability to recognize words, which causes difficulties in writing and understanding text. Despite normal levels of intelligence and adequate teaching, children with dyslexia often have difficulty acquiring reading skills at the level of their peers [2]. However, the effects of dyslexia are not limited to the classroom, but can also lead to problems with self-esteem, frustration, and social isolation [3,4,5]. Dyslexia is a complex disorder that affects reading, writing, and language processing abilities. Although it is most commonly diagnosed in school-aged children, its causes are varied and may include both neurological and genetic factors. Among the hypotheses regarding the causal background of dyslexia, the following are suggested:

• Phonological disorders related to difficulties in processing speech sounds, such as phonemes—the basic sound units of language—are the primary cause of dyslexia. Children with dyslexia struggle with segmenting, blending, and manipulating sounds in words, which directly impacts their reading and writing skills. Research confirms that deficits in phoneme recognition affect reading and writing abilities [2,6].
• Problems with the coordination and automation of cognitive processes related to the cerebellum can lead to difficulties in learning to read, as confirmed by neuroimaging studies [7].
• Deficits in the magnocellular cells responsible for processing rapid visual stimuli can cause difficulties in reading fluency and visual perception [8].
• Improper migration of neurons during brain development can lead to dysfunction in areas responsible for language processing and reading [8].

The symptoms of dyslexia include difficulties in word recognition, spelling problems, slow reading speed, and challenges in segmenting speech sounds. Current methods for diagnosing dyslexia involve reading skills tests, phonological assessments, psychological tests, and interviews with parents and teachers.

The study of dyslexia is crucial due to its significant impact on educational outcomes and social functioning. Dyslexia can affect a child’s ability to read and write effectively, leading to broader educational challenges, including lower academic achievements and an increased risk of early school dropout [2]. The economic burden associated with dyslexia is considerable in terms of educational costs, therapy, and the potential loss of productivity due to the challenges faced by individuals with dyslexia.

Dyslexia is a serious issue, not only for individuals but also for society as a whole. It is one of the most common learning disorders, affecting a significant portion of the population. Dyslexia impacts about 5–10% of the global population [9]. In the United States, the National Institutes of Health (NIH) estimate that around 15–20% of the population exhibits symptoms of dyslexia [3]. In Europe, a comprehensive review conducted by the European Dyslexia Association highlights that dyslexia affects about 10% of the population, with varying degrees of severity [10]. These figures underscore the widespread prevalence of the disorder and its impact on educational systems and individual lives.

The standard approach to treating dyslexia does not involve using a drug or an established therapy for all patients, as is the case with treating a specific disease. Dyslexia is not a disease but a developmental disorder that manifests differently in different individuals. Each of us is unique, and depending on the individual characteristics of our bodies, we experience this disorder in various ways. The research presented in this paper responds to a new approach to the diagnostic–therapeutic process, changing the paradigm of therapy and rehabilitation for neurodevelopmental disorders, especially in children and adolescents. The proposed technology for diagnosing dyslexia departs from the traditional approach that applies the same treatment method to all individuals with the same disorder. People differ from each other, and therefore, there is no universal method of therapy.

The last few decades have brought about significant progress in understanding and diagnosing dyslexia, a disorder characterized by difficulties in reading. Research on dyslexia has also shown that it is not a homogeneous disorder but rather a spectrum of difficulties manifesting in various ways, necessitating an individualized approach to diagnosis and therapy. This individualized approach ensures that each person’s unique needs are addressed effectively. As highlighted in [3,11], dyslexia is a developmental disorder that requires individualized and adaptive treatment strategies to effectively address the unique needs of each person. Among the methods used to diagnose dyslexia, particular attention has been given to those that allow for the direct observation of behaviors and eye movements during reading [11,12]. Dyslexia, traditionally diagnosed based on assessments of reading and writing skills, has gained new diagnostic tools thanks to the development of digital technologies [6,13,14,15]. The digitization of tests allows for a detailed analysis of the reading process, providing data that are difficult to obtain in traditional settings [16,17]. Traditional assessment methods, focusing mainly on written and oral tests, are increasingly supplemented by modern technologies such as eye tracking and artificial intelligence, offering new possibilities for objective and precise diagnostics. Diagnostic methods based on brain imaging, such as functional magnetic resonance imaging (fMRI) or diffusion tensor imaging (DTI), provide valuable information about structural and functional differences in the brains of individuals with dyslexia compared to control groups [18,19]. These techniques, supplemented by electroencephalographic (EEG) studies, shed light on the neurobiological basis of dyslexia, explaining the mechanisms responsible for reading difficulties [20,21,22].

Eye tracking, or tracking eye movements, has become one of the key tools in research on dyslexia. It allows for precise tracking of how individuals with dyslexia read text, including the analysis of eye fixation patterns, saccades (rapid eye movements), and other eye movement characteristics that may indicate difficulties in text processing [23,24,25]. Innovative studies, such as [26], emphasize the importance of eye movement coordination in children with dyslexia, showing that they may have problems with proper text tracking, directly impacting their reading skills. Conversely, in the work of [27], artificial intelligence is used to predict dyslexia based on reading patterns in children, demonstrating how modern technologies can revolutionize the diagnosis of this disorder. In the work of [28], attention is drawn to the potential of using machine learning and eye tracking to identify individuals with dyslexia, opening new perspectives for precise and rapid diagnosis. Similarly, further research with larger sample sizes and advanced data analysis methods presented in [29,30] demonstrate that the application of sophisticated algorithms can significantly improve the accuracy of dyslexia detection, achieving an effectiveness level of over 95%. The development of deep learning algorithms and their application in eye tracking data analysis opens new possibilities in dyslexia diagnosis. In the works of [31,32], utilizing neural networks for eye tracking data processing, the potential of these technologies for identifying reading disorders is highlighted, offering high efficacy and paving the way for the development of new, even more efficient diagnostic tools.

Table 1. The review of recent findings in eye tracking for dyslexia diagnosis.

Authors	Subject	Age	Experimental Approach	Main Findings
Christoforou, C.; et al. [23]	30 children with dyslexia and 30 chronological age controls	A mean age of 9.79 years and a range of 7.6 to 12.1 years.	A combined EEG and eye- tracking study on children with dyslexia	Novel framework for integrative analysis of neurophysiological and eye-gaze.
Jakovljevi, T.; et al. [24]	36 children, 18 with dyslexia and 18 control	8–12 years old	The reading task in 13 combinations of background and overlay colours	Findings showed that the dyslexic children have longer reading duration, fixation count, fixation duration average and total, and longer saccade while reading on white and coloured background/overlay.
Jakovljevi, T.; et al. [25]	25 children, 10 boys and 15 girls	8–9 years	This study investigated the influence of white vs. 12 background and overlay colors on the reading process.	The findings showed a decreasing trend with age regarding EEG power bands and lower scores of reading duration and eye-tracking measures in younger children compared to older children.
Temelturk, R.D.; et al. [26]	Children with dyslexia and typical development	5–17 years	The review through the examination of binocular coordination in children with dyslexia by describing the normative development of stable binocular control.	The studies reviewed provided consistent evidence of poor binocular coordination in children with dyslexia.
Wang, R.; et al. [27]	399, 187 with dyslexia and 212 typically developing children	7–13 years	These studies implemented tests evaluating reading-related cognitive skills.	This study established a genetic algorithm optimized back-propagation neural network model to predict whether Chinese children have dyslexia.
Rello, L.; et al. [28]	97, 49 without dyslexia (28 female, 21 male) and 48 with diagnosed dyslexia (22 female, 26 male)	11–54 years	Each participant read 12 different texts with 12 different typefaces. The texts and the fonts were counter balanced to avoid sequence effects.	The eye movements of readers with dyslexia are different from regular readers. People with dyslexia have longer reading times, make longer fixations, and make more fixations than readers without dyslexia.
M. N. Benfatto, et al. [29]	185, 97 high-risk subjects and a control group of 88 low-risk subjects.	9–10 years	Using eye tracking during reading to probe the processes that underlie reading ability.	It is possible to identify 9–10-year-old individuals at risk of persistent reading difficulties by using eye tracking.
Prabha, A.J.; et al. [30]	185, 97 high-risk subjects, a control group of 88 low-risk subjects.	9–10 years	Using eye tracking during reading to probe the processes that underlie reading ability.	The research focused on identifying features that contribute to better prediction and then build an appropriate prediction model.
Neruil, B.; et al. [31]	185, 88 with low risk (69 male,19 female) and 97 with high risk of dyslexia (76 male, 21 female)	9–10 years	A new detection method for cognitive impairments is presented utilizing eye tracking signals in a text reading test.	In a series of experiments it was found that the best results provide magnitude spectrum-based representation of the time-interpolated eye-tracking signals recorded.
Vajs, I.; et al. [32]	30 persons (19 female, 11 male), 15 with dyslexia and 15 control subjects.	7–13 years	The children read a text written in Serbian on 13 different color configurations (including background and overlay color variations).	A combination of convolutional neural network and visual encoding of the eye tracking data shows promising results in dyslexia detection with minimal preprocessing effort.

presents the main directions of existing research that utilize eye-tracking methods in diagnosing dyslexia [23,24,25,26,27,28,29,30,31,32].

In summary, the dynamic development of research on dyslexia, supported by technological progress, is changing the face of diagnosing this disorder. The use of modern tools such as eye tracking and machine learning not only increases the precision of dyslexia recognition but also helps to better understand the underlying mechanisms, opening new pathways for effective therapeutic intervention and support for individuals with this disorder.

In the research presented in [33,34,35], the subsequent part of the study builds upon previously conducted hybrid studies in the field of developmental psychology, which involved therapeutic and preventive interventions among children aged 10–14 years old. The author’s program consisted of selected elements from various therapeutic interventions, particularly Davis, CBT, SI, hand therapy, and eye training. The aim of the research was to alleviate difficulties in writing/dysgraphia. As a result of the conducted program, participants were observed to exhibit correct muscle tension in the fingers and wrist, proper writing grip, and a correct habit associated with writing technique. The therapeutic interventions were conducted in face-to-face settings. Through the conducted research, a strategy for designing and utilizing/conducting psycho-tests and utilizing attention analysis to assess the effectiveness of therapy and individual therapy selection was developed. Additionally, over the course of several years, the team conducted studies on the attention of pilots during the execution of specific types of aviation tasks in both IFR and VFR conditions, demonstrating that the observer’s attention is diffuse, while the shape and trajectory of attention over time–space are indicators of the pilot’s training level. It was shown that the chronology of attention is directly linked to the pilot’s ability to perceive information from cockpit instruments and directly impacts flight safety [36,37]. It was proven that the shape and dynamics of observation trajectories are directly related to the process of scene recognition and the perception of its individual components. Therefore, there is a strong coincidence between the understanding of a scene and the dynamics of observation, which can serve as a significant source of diagnostic data for a hypothetical neural system. The authors thus combined their experience using human–machine interface (HMI) systems with the new capabilities of recurrent neural networks (LSTM), which are currently an efficient and effective tool for analyzing and recognizing time series [38,39]. Based on this groundwork, the present study aims to develop intelligent technology that will support therapy for neurodevelopmental disorders. The proposed combination of research related to psycho-tests and the measurement of the attention of the subject will allow for the establishment of individualized therapeutic procedures for each patient, minimizing adverse effects.

The proposed hybrid technology for conducting psycho-tests, allowing for result assessment through the use of observer attention, represents an original and innovative approach, adding value to this study. In this article, we propose an innovative approach to diagnosing dyslexia by integrating the Benton Visual Retention Test (BVRT) with advanced analysis of visual attention trajectories using eye-tracking technology and deep neural networks (DNN). This approach not only enhances the effectiveness of dyslexia diagnosis but also makes it more accessible and efficient, especially for children who are not yet able to read. The Benton test is a widely recognized tool for assessing visual memory and perception. Integrating this test with eye-tracking technology allows for detailed analysis of a child’s eye movement patterns while performing tasks. Recording the spatiotemporal trajectories of visual attention using mobile devices like Pupil Invisible or Pupil Core provides objective and precise data. The recorded spatiotemporal data, extracted from sessions, are analyzed by an LSTM network, enabling the detection of subtle visual anomalies characteristic of dyslexia. One of the main advantages of the proposed approach is its ability to diagnose dyslexia in children who have not yet mastered reading skills. Traditional diagnostic methods often rely on reading tests, which can be a barrier for younger children or those with severe reading difficulties. The method based on visual perception and eye movement analysis bypasses this obstacle, allowing for the early detection of dyslexia. The proposed solution offers several key advantages over traditional dyslexia diagnostic methods:

• Objectivity and Precision: Utilizing eye tracking for the accurate and objective collection of data related to eye movements and subjecting the gathered data to analysis using LSTM networks allows for the identification of subtle eye movement patterns characteristic of dyslexia. The high correlation coefficient R (∼0.992) achieved in the proposed model indicates its high accuracy and reliability.
• Speed and Efficiency: Traditional diagnostic methods can be time-consuming and require multiple sessions with the child. The proposed approach allows for rapid data collection and immediate analysis of results, significantly reducing the time needed for diagnosis.
• Stress-Free Environment: The BVRT, which does not require reading skills, is less stressful and more natural for younger children. Eye tracking allows for administering the test in a friendly and engaging manner, which can lead to more reliable results.
• Early Intervention: The ability to diagnose dyslexia in children who are not yet able to read enables the early implementation of appropriate educational and therapeutic interventions. Early recognition of dyslexia-related issues allows for the prompt introduction of effective support strategies, which can significantly improve the child’s educational and emotional outcomes.

A key added value of this work is the combination of the Benton Visual Retention Test with eye-tracking technology and LSTM deep neural networks, presenting a novel approach to diagnosing dyslexia. The ease of implementing this approach in school settings and its stress-free nature make it suitable for use with children of various ages and developmental stages, including those who have not yet learned to read. This enables early intervention, which is crucial for effective educational and emotional support for children with dyslexia.

2. Materials and Methods

2.1. Participants

The study included 9-year-old children attending the third grade of primary school in Jarosław, southeastern Poland. All participants were primary school students from one geographical region, allowing for a certain level of socioeconomic and educational homogeneity. A total of 145 children participated in the study (66 girls and 79 boys). In 65% of the cases, at least one parent had higher education, 20% had a secondary school education, and 15% of parents had a primary school education. In the studied group, 40% of the children came from families with a middle socioeconomic status, 35% from families with a high status, and 25% from families with a low status. A total of 90% of the children lived in urban areas, while 10% lived in rural areas. In the studied group, a diagnosis of dyslexia was made at the age of 8, following earlier observations (and a diagnosis in the “zero” class—at risk of dyslexia) indicating a risk of dyslexia between the ages of 5 and 7. All children participated in therapy lasting an average of 1.5 years, which included various support methods, including speech therapy and psycho-pedagogical therapy. Some children had coexisting visual impairments that could have affected their performance in visual studies. However, all children underwent appropriate ophthalmological diagnostics and, if necessary, treatment before starting the dyslexia study. The study was approved by the Institutional Review Ethic Board of the PGKICPO, and the ethical approval was granted on 9 November 2022, under the reference number IRB-20221109. Informed consent was obtained from all subjects involved in the study, and written informed consent has been obtained from the patient(s) to publish this paper.

After assembling the group of participants for the experiment, research was conducted on a test group comprising 145 individuals aged 7 to 10 years old. The experiment for each participant lasted no longer than 10 min, with the duration depending on the individual and the time allocated for reproducing the pattern from memory. The summary of the conducted experiments is presented in

Table 2. Characteristics of the subjects.

Name	Value
Number of people surveyed	145
Number of women	66 (46%)
Number of men	79 (54%)
Right-handed persons	132 (91%)
Left-handed persons	13 (9%)
People with visual impairment	7 (5%)
Individuals without a visual defect	138 (95%)
Maximum duration of the study	00:08:32
Minimum duration of the study	00:03:37
Average duration of the study	00:04:54

.

2.2. Research Equipment

For the purposes of the conducted research, a research station was designed and configured, enabling the registration of observer attention. The data collected in this way allow for the processing of the obtained video sequence using DNN, thereby enabling the diagnosis of neurodevelopmental disorders associated with dyslexia. Figure 1 illustrates the utilized research station.

The research was conducted using two eye-tracking systems: Pupil Core v2.0.182 and Pupil Invisible. These systems offer high precision in tracking eye movements, which is crucial for their use in experimental research, human–computer interfaces, and virtual reality. The systems are equipped with dedicated cameras for recording eye movements and a camera for recording the observed scene. Pupil Core, in its initial phase, requires calibration, which allows for precise adjustment of eye movement tracking to a specific user. To ensure high measurement accuracy, a 5-point calibration and natural calibration using Apriltag markers were utilized. Pupil Invisible is an attention-tracking system based on deep learning, which eliminates the need for calibration and significantly increases measurement reliability.

During the measurements, the least-invasive measurement model was adopted, which utilized the Pupil Invisible eye tracker. The recorded video stream for each participant was processed in the Pupil Cloud. In this way, measurement data were obtained, which were used to create a training set. An example of recording observer fixations during the object reproduction task from the Benton test is presented in

Table 3. Examples of fixation data obtained from eye-tracking measurement.

Fixation id	Duration [ms]	Fixation Detected on Surface	Fixation x [Normalized]	Fixation y [Normalized]
2	350	True	0.381184	0.434751
3	155	True	0.478635	0.452984
4	216	True	0.351071	0.473251
5	79	True	0.164494	0.872142
6	735	True	0.148455	0.874538
7	351	True	0.055462	0.814655
8	371	True	0.166153	0.897663
9	991	True	0.133664	0.877281
10	135	True	0.27131	0.96255
11	220	True	0.377649	0.961936
12	223	True	0.43152	0.615011
13	323	True	0.359779	0.514438
14	251	True	0.740871	0.85324
15	131	True	0.497466	0.389931
16	199	True	0.229441	0.346479
17	152	True	0.173246	0.405854
18	231	True	0.37687	0.408425
19	368	True	0.543391	0.358734
20	167	True	0.468531	0.36243
21	251	True	0.537559	0.377421
22	531	True	0.427361	0.363313
23	91	True	0.409878	0.371664
24	1331	True	0.375001	0.397355
25	240	True	0.502381	0.388864
26	300	True	0.596782	0.367004
⋮	⋮	⋮	⋮	⋮
40	259	True	0.602571	0.321279
41	326	True	0.423158	0.297279
42	319	True	0.995357	0.347378
43	751	True	1.01715	0.383466
44	256	True	0.478577	0.393537
45	159	True	0.48118	0.368611
46	60	True	0.340173	0.727317
47	112	True	0.378526	0.787224
48	291	True	0.001522	0.751218
49	92	True	0.262708	0.446477
50	676	False	1.133303	0.247477
51	375	False	1.120387	0.232185
52	223	False	1.089224	0.161522
53	160	True	0.47592	0.410375
54	168	True	0.80732	0.340732
55	156	True	0.922784	0.25163
56	188	True	0.949428	0.552949
57	208	True	0.921888	0.298959
58	68	True	0.495532	0.261822
59	91	True	0.546163	0.258226
60	227	True	0.585588	0.310388
61	191	True	0.27676	0.302838
62	223	True	0.186211	0.256423
63	156	True	0.324427	0.296805
64	160	True	0.324148	0.17112
65	415	True	0.43132	0.149982
66	136	True	0.285724	0.3394
67	668	True	0.270591	0.162541

. Fixations marked as True indicate that the object of attention is within the observed scene plane defined by the Apriltag set. The value False indicates that the observer’s attention momentarily moved outside the field of view. Figure 2 shows a graph for the recorded data obtained during the BVRT survey from Table 3. This data in the form of a time moving window were the input to the LSTM network. In the experiments presented in this work, a sampling frequency of 120 Hz was adopted. The data stream recorded during the experiments was processed in the Pupil Cloud environment. This resulted in observer attention trajectories with a non-uniform time axis, chronologically encompassing the moments of fixation occurrence, their durations, and the coordinates of the detected fixations normalized relative to the adopted coordinates of the observed scene. The resulting non-uniform time series of fixation coordinates was processed using a window size of 256 with a shift step of 2, which considered the lengths of the recorded trajectories in all experiments and allowed for proper balancing of the training dataset. The observation window size was selected heuristically to account for both the lengths of the processed trajectories (the shortest being 387 fixations, and the longest being 574 fixations) and to maximize resistance to temporary measurement disturbances (occlusions, blinks, going outside the controlled field of observation, etc.).

2.3. Description of the Task to Be Carried Out

During the execution of individual studies, participants were asked to perform the Benton Visual Retention Test, assessing visual memory. BVRT allows for inferences about potential changes in the overall neurological status of the patient based on obtained results, especially regarding visual perception, visual memory, and visuoconstructive abilities. It is a sensitive diagnostic tool used, among others, in reading difficulties, traumatic brain injuries, and attention deficit disorders. BVRT has three alternative forms: C, D, and E, all of which are equivalent and can be administered in different conditions. In the conducted experiments, version C of the test was used, along with method A: exposure of the pattern for 10 s followed by immediate reproduction from memory. The test material consisted of geometric figures placed on a white background. Views of the test cards are presented in Figure 3.

Table 4. Observed dyslexia levels.

Level of Visual Perception Disorder	Range	Quantity	%
Low	1–5	84	58
Average	6–7	35	24
High	8>	26	18

classifies the relevant risk levels of visual perception disorders, as indicated by the range of errors made by participants. Participants were classified based on the number of errors made and the overall accuracy in reproducing the patterns.

The correctness of copying patterns as well as their reproduction from memory were assessed, considering both the number of correct drawings and the number of errors made. Errors indicating spatial function disturbances involve omitting, distorting, rotating, or repeating (perseverating) memorized figures from the previous pattern.

The visual perception experiment proceeded according to the following steps:

• Taking a seat at the research station, ensuring appropriate measurement conditions, in a position similar to that typically assumed by participants when seated at a desk during lessons or other activities. This ensured the naturalness of the research environment;
• Wearing the necessary glasses for conducting the visual perception test. (In the case of using the Pupil Core system, a calibration process was conducted);
• Familiarizing oneself with the instructions regarding task execution;
• Testing the interactive tool for reproducing pattern exposure;
• Performing the BVRT test face-to-face, consisting of 10 cards (pattern exposure followed by immediate reproduction from memory).

In Figure 4, selected stages from the conducted experiments are presented. In Figure 5, sample shots from the participant’s world camera and information recorded by the eye tracker during the experiments are shown.

In the initial phase of the experiments, the team conducted 10 preliminary measurements aimed at demonstrating the existence of distinct features distinguishing attention trajectories depending on the degree of dyslexia risk. Subsequently, a classical assessment of participants’ work sheets was performed, determining the dyslexia risk level (DRL) coefficient. The visualization of attention trajectories depicted in Figure 6 indicates that the dispersion of attention for individuals with a high risk of dyslexia is minimal, whereas for those with a low risk, it is greater. This suggests, on the one hand, greater perceptual mobility in healthy individuals and, on the other hand, the existence of characteristic features that can be utilized in the neural network learning process. Preliminary trajectory sets, whose observation durations indicate that healthy individuals focus their attention on the observed scene for a longer period, thus effectively completing the reproduction task, are listed in

Table 5. Dyslexia risk level and trajectory length coincidence, the best and the worst case marked with bold respectively.

Registered Attention Trajectory	Dyslexia Risk Level
2 × 461 double	0.4
2 × 574 double	0.3
2 × 506 double	0.2
2 × 637 double	0.4
2 × 576 double	0.4
2 × 548 double	0.1
2 × 482 double	0.1
2 × 357 double	0.6
2 × 477 double	0.6
2 × 387 double	0.8

.

Some of the strands of research that have formed in existing methods of detecting dyslexia with AI methods include [13], in which the image analysis of reproduced BVRT test cards was used, but it did not take into account the chronology and dynamics of scene observation or the relationship between observation and drawing tool operation. The coincidence of fixations depending on the complexity of the observed geometric figure is depicted in Figure 7.

During task execution, additional parameters of the process are recorded, such as momentary pupil diameter, which may be associated with individual characteristics of the study participant (see Figure 8).

In the next chapter, based on the preliminary research conducted and observations, and utilizing the accumulated experience of team members, a measurement setup using a DNN (deep neural network) was proposed.

3. Attention-Tracking System Utilizing DNN to Support Therapy for Individuals with Neurodevelopmental Disorders

In Figure 9, a modular structure of an attention-tracking system supporting therapy for individuals with neurodevelopmental disorders is presented. Module A encompasses a designed research setup for recording the observer’s attention during the performance of the BVRT test. This module is responsible for acquiring the video stream of the observed scene and recording the spatial position of the observer’s pupils. Module B performs feature extraction (1) related to attention coordinate detection and fixation detection. The obtained data are passed to component (2), where tasks associated with recognizing visual perception disorders using LSTM (long short-term memory) networks are performed. The obtained results are transmitted to module C, and based on them, the therapist decides whether to initiate therapy or refer the subject to additional sessions. In contrast to previous works, which mainly focused on the use of neural networks during the reading and writing process [7,40,41,42,43,44], this study utilizes an LSTM network in the process of diagnosing dyslexia in elementary school-aged children. For implementation purposes, the Matlab environment and dedicated libraries for modeling DNNs were utilized. From the literature, it is known that attention trajectories, which possess a time-series nature, are effectively processed by LSTM networks. Therefore, the authors adopted a 5-layer model of the network, the structure of which is included in

Table 6. Net structure from the deep learning network analyzer.

Name	Type	Activations	Learnables
Trajectory series	Sequence input	5		–
Attention analyser	LSTM	16	InputWeights	64 × 12
			RecurrentWeights	64 × 16
			In Bias	64 × 1
Encoder	Fully Connected	6	Weights	6 × 16
			Bias	6 × 1
Softmax normalizer	Softmax	6	–
Dyslexia clasifier	Classification Output	6	–

.

Within the LSTM layer, the input gates x and y were adopted, respectively, which process successive fixations in the attention trajectory that occur during the measurement process. In the diagram below, $h_t$ denotes the output, also referred to as the hidden state, while $c_t$ denotes its state at time step t (see Figure 10).

At the time step t, the consecutive net layer cells use the current state of the RNN $\left(c_{t-1},h_{t-1}\right)$ and the next time step of the sequence to compute the output and the updated cell state $c_t$. The hidden state at time step t contains the output of the LSTM layer for this time step. The cell state contains information learned from the previous time steps. Each cell controls updates using gates (see Figure 11).

The diagram shows how the gates forget, update, and output generate the cell signal and its hidden states, respectively. The learnable weights of an LSTM layer are

$$\begin{array}{c}\hfill W={\left[W_i{W}_f{W}_g{W}_o\right]}^T\hfill \end{array}$$

(1)

$$\begin{array}{c}\hfill R={\left[R_i{R}_f{R}_g{R}_o\right]}^T\hfill \end{array}$$

(2)

$$\begin{array}{c}\hfill b={\left[b_i{b}_f{b}_g{b}_o\right]}^T\hfill \end{array}$$

(3)

where W denotes input weights, R represents recurrent weights, and b indicates bias, respectively. The flow of signal through the net implies that matrices W, R, and b are concatenations of the input weights, the recurrent weights, and the bias of each component, respectively. Additionally i, f, g, and o denote the input gate, forget gate, cell candidate, and output gate, respectively. The cell state at time step t is given by

$$\begin{array}{c}\hfill c_t=f_t\odot c_{t-1}+i_t\odot g_t\hfill \end{array}$$

(4)

where ⊙ denotes the Hadamard product. The hidden state at time step t is given by

$$\begin{array}{c}\hfill h_t=o_t\odot {\sigma }_c\left(c_t\right)\hfill \end{array}$$

(5)

where ${\sigma }_g$ denotes the state activation function. We assumed default transfer function as the hyperbolic tangent function to compute the net state at consecutive states [45].

One of the key stages in preparing data for analysis was standardizing the duration of fixations. Due to the large variation in the values of this feature, its standardization was applied, allowing for comparison and analysis of visual behaviors among participants. This stage was necessary to adapt the data for further modeling, considering the diversity of natural visual behaviors in children.

4. Results and Discussion

The analysis of the Benton Visual Retention Test (BVRT) results conducted on a group of early school-age children provided valuable insights into their cognitive abilities and potential learning difficulties, particularly related to dyslexia. The study involved 145 children, each subjected to a series of trials aimed at assessing their ability to reproduce geometric patterns from memory. Based on the obtained results, appropriate risk levels of visual perception disorders were classified, as indicated by the range of errors made (see Table 4). Participants were classified based on the number of errors made and the overall accuracy in reproducing the patterns.

The total number of errors made by the study participants consists of various types of inaccuracies, indicating the complexity of cognitive processes related to visual perception, visual memory, and visual constructional abilities. A statistical summary of the number of correct and incorrect reproductions is presented in

Table 7. Summary of errors.

Name	Quantity
Number of correct mappings	746
Number of incorrect mappings	704
Average number of correct mappings	5.14
Average number of misrepresentations	4.86

.

According to the task assumptions, the correctness of pattern execution and its reproduction from memory were evaluated. Errors indicating spatial function disturbances were mainly related to distortions, omissions, and displacements. A detailed breakdown of errors is presented in

Table 8. Summary of error types.

Types of Errors in Subjects	Quantity
Skip	190
Distortion	212
Perseverations	61
Rotation	58
Translation	139
Errors of relative magnitude	44

.

Analyzing the results obtained from the conducted eye-tracking studies, differences between fixations during memorization and reproduction of figures can be observed. As shown in Figure 7, trajectories during observation and reproduction differ slightly. This can be particularly observed in the attention plots during memorization and reproduction of shapes. The Figure 12 illustrates the convergent learning process for the adopted network architecture. As can be seen, this process is highly unstable due to numerous similarities among individual attention trajectories. Additionally, one must consider the arbitrary manner in which the expert assigns scores in the BVRT test, which generates a locally biased information leakage effect.

For the network used, a fit test of the trained network model was carried out using a linear regression model. The following results, shown in

Table 9. Summary of error types.

	Estimate	SE	tStat	p Value
(Intercept)	0.03301	0.052743	0.62586	0.53488
×1	0.99709	0.014096	70.734	1.7749 × ${10}^{-44}$

Number of observations: 43, Error degrees of freedom: 41. Root Mean Squared Error: 0.154. R–squared: 0.992, Adjusted R–Squared: 0.992. F–statistic vs. constant model: 5 × ${10}^3$, p-value = 1.77 × ${10}^{-4}$.

, were obtained for the designed network during testing mode (train/test ratio: 70%/30%). In Figure 13, a confusion matrix is also presented, which facilitates the final assessment of the determination level of the obtained model. In the test set containing 43 records, 42 records were identified correctly with one incorrect diagnosis, yielding an overall prediction accuracy of the LSTM network of 97.7%, which should be considered a high indicator.

Based on this, it can be concluded that the obtained coefficient of determination $R^2$ indicates the high effectiveness of the proposed method of attentional analysis for diagnosing dyslexia.

Early diagnosis of dyslexia is essential for maximizing the potential of individuals with dyslexia and for creating a more inclusive, educated, and economically stable society. By identifying and addressing dyslexia early, we can ensure that all individuals have the opportunity to succeed and contribute positively to their communities. Early diagnosis of dyslexia is especially essential for providing children with the support they need to succeed academically, emotionally, and socially. It empowers families, educators, and society to create a more inclusive and effective education system, leading to better outcomes for individuals and communities.

Over the past few years, various tools and methods have been developed to achieve relatively high detectability of dyslexia. These include studies of brain activity associated with cognitive processes during specific tasks using highly specialized equipment, such as fMRI imaging. Analysis of these images with convolutional networks has achieved an accuracy of $72.73\%$ [13,19]. Combining fMRI with DTI can provide extended DICOM data, which, when analyzed using PCA, serve as a source of information for neural classifiers, achieving an accuracy of $94.87\%$ [18].

In contrast, other systems [13] (with an accuracy of around $94.73\%$) are convenient and easy to use, as they require only modules for acquiring graphical task results. This is an important feature that facilitates their application in school settings without the need for an expert’s involvement. On the other hand, the age and academic skills of the children being tested may necessitate differentiating the test tasks used for dyslexia detection.

Methods for diagnosing dyslexia that utilize reading process analysis often employ SVM binary classifiers [28], which leverage the properties of fixation and saccade observations, achieving an accuracy of $80.18\%$. Higher accuracy is achieved by KNN classifiers [30], tested on a relatively large group of children, reaching an accuracy of $95\%$.

Some reading task analysis methods are tailored to specific languages, such as Serbian [32], where analysis of the geometry of observation trajectories achieves an accuracy of $87\%$. In methods aimed at individuals with reading and writing skills, both lexicographic and semantic tasks can be used. Efforts to increase dyslexia detection accuracy also focus on acquiring a large number of coordinates representing the state of the perception and scene recognition process. In such cases, the task of extracting relevant information is left to CNNs, achieving an accuracy of $95.6\%$ [31].

5. Conclusions

This study presented significant findings on using deep LSTM networks for dyslexia recognition by analyzing time series data depicting the attention trajectories of individual participants. Among various eye trackers, including Tobii, SMI, and Pupil, the team effectively utilized the Pupil Invisible and Pupil Core models, which facilitated seamless research in school environments.

Based on the conducted experiments, the following conclusions can be drawn:

• The Benton Test, employed to establish the expected values of the training set, proved to be an effective tool for use in intelligent systems aimed at recognizing developmental disorders in early school-age children.
• For dyslexia detection studies in children, non-invasive eye trackers that minimally disrupt the child’s attention during measurements are recommended.
• Spatiotemporal measurements of attention trajectories can be effectively utilized to identify anomalies indicative of dyslexia risk.
• A high level of dispersion in attention trajectories correlates with high accuracy in task reproduction during BVRT tests, suggesting a lower risk of dyslexia.
• As with other systems that rely on expert knowledge, the arbitrary assessment method of BVRT test results for constructing a learning sequence is a limitation. It is advisable to involve a larger number of experts for independent result assessments to enhance reliability.
• The definite ease of implementation in a school setting (preferring a Pupil Invisible stand) of the proposed method is a significant advantage and superiority over methods using fMRI measurements [7].
• Analyzing attention trajectories using LSTM networks offers a robust alternative to methods utilizing CNNs for static graphical analysis of the BVRT test forms, as it accounts for the temporal and spatial strategies employed by humans in scene recognition.

Based on the observations and conclusions from the conducted research, several promising directions for further studies on dyslexia recognition can be identified. In particular, it is anticipated that combining LSTM and CNN networks in future research will enhance the accuracy and confidence in dyslexia recognition results for early school-age children. Exploring the creation of a hybrid model that combines various methods of analyzing perception and scene recognition, as well as psychomotor reactions during performing specific tasks, will be a focus for the authors’ future work.