ETDD70: Eye-Tracking Dyslexia Dataset

The ETDD70 dataset comprises eye-tracking recordings from 70 Czech participants, equally divided into dyslexic and non-dyslexic readers, all aged 9–10 years. The dataset captures eye movements during three text-reading tasks in Czech: syllable reading (Task 1), meaningful text reading (Task 4), and pseudo-text reading (Task 5).

This dataset is the result of the project “Diagnostics of Dyslexia Using Eye-Tracking and Artificial Intelligence” conducted by our research team. The project aims to leverage artificial intelligence tools and advanced technical equipment (eye tracking) to more effectively diagnose dyslexia, one of the most common specific learning disorders, and thereby significantly improve re-education strategies for dyslexic students. The primary goal is to develop models that accurately distinguish between dyslexic and non-dyslexic readers based on eye movement patterns recorded during these tasks.

Data collection took place between October 2022 and August 2023, adhering to ethical standards. The project was approved by the Research Ethics Committee of Masaryk University in Brno, Czech Republic.

Please contact us if you have any questions or feedback at [e-mail] or at [e-mail].

The ETDD70 dataset is freely available for research purposes. We kindly ask users to contact us (at [e-mail]) prior to any research use, so we can keep track of applications and provide relevant guidance.

PARTICIPANTS

The eye-tracking data were captured from 70 participants: 35 dyslexic and 35 non-dyslexic readers. In all cases, participants are elementary school pupils aged 9-10 years (i.e., 4th grade of elementary school). Recruitment of suitable participants was conducted in cooperation with a psychological counseling center, which facilitated the recruitment of pupils diagnosed with dyslexia. The non-dyslexic readers, who showed no symptoms of dyslexia, were recruited in cooperation with the counseling facilities of selected elementary schools. The dataset was collected from October 2022 to August 2023. The legal representatives of all participants were properly informed about the research procedure and agreed to participate in the study, for which they subsequently received compensation.

STIMULI

We designed three verbal tasks based on standardized paper-based dyslexia diagnostics used in the Czech Republic. These source texts were transferred to a digital version in a controlled form (e.g., amount of text, font size, line spacing, background color, etc.) for the requirements of eye-tracking measurements.

Task called Syllables contains 90 syllables arranged in a 9 x 10 matrix. The syllables are commonly encountered in the Czech language. The individual rows of syllables were categorized according to syllable composition (based on linguistic aspects) as follows: open syllables with no meaning, i.e., consonant + vowel (e.g., "ta," "na"), closed syllables with a central vowel bearing a meaning, i.e., consonant + vowel + consonant (e.g., "suk," "mák"), meaningless syllables consisting of two consonants (e.g., "vl," "bz"), a meaningless syllable formed by a cluster of two consonants ending in a vowel (e.g., "tle," "mra"), and finally a meaningful syllable formed by a cluster of three consonants with one vowel in the 3rd position (e.g., "mrak," "vlak"). All syllables were presented in black font, with Times New Roman on a gray background. The objective of the task is to read aloud all syllables from left to right and from top to bottom. A fixation cross was placed in the lower right corner for gaze-contingent task closure—when the participant looks at this cross, the recording is automatically terminated.

Task called MeaningfulText consists of a passage about a young boy who watches a squirrel from his window. This text is intended for elementary school readers in grades 3 and 4. The stimulus text contains a total of seven text lines with six logical sentences. The text is again written in black-colored font with double line spacing on a grey background and the fixation cross in the lower right corner. The aim of the task is to read the entire text aloud.

Task called PseudoText comprises fictional, meaningless words. This text has a total of seven lines with 15 artificial sentences. The text formatting, as well as the ending fixation cross, are the same as in Task  MeaningfulText. The objective of the task is to read the entire text aloud as smoothly as possible.

EYE-TRACKING FEATURES

The raw eye-tracking data recorded for each task were further processed to extract event-based characteristics—fixations, saccades, and dozens of derived statistical characteristics. The fixations were detected using the i2mc algorithm (Hessels et al., 2017), as it was specifically designed to be noise-robust for measurements in children. The minimum fixation duration was set to 40 ms. The derived characteristics provide additional information about how participants interact with text. These characteristics are divided into whole-task and region-of-interest (ROI) characteristics. While the whole-task characteristics describe the semantics on the global level of the whole screen, the ROI ones characterize the semantics on the local level of a small rectangular area.

Feature-based characteristics for each task:

Syllables

First fixation duration, average fixation duration, number of fixations, number of fixations and saccades without the incoming/outgoing saccade, number of revisits—incoming saccades hitting this ROI from outside.

MeaningfulText, PseudoText

Whole-task (features extracted from the whole trial): number of regressions, ratio of progressive to regressive saccades, average saccadic amplitude, total reading duration, average fixation duration, number of fixations.

ROI (features extracted for separated regions of interest, i.e. lines and words): average fixation duration, number of fixations, number of revisits—incoming saccades hitting this ROI from outside, landing position of the first fixation.

AI CLASSIFICATION APPROACH

The AI-based methods used for the classification of dyslexia are available here: https://gitlab.fi.muni.cz/xsedmid/dyslex

CITE THIS DATASET

Dostalova, N., Svaricek, R., Sedmidubsky, J., Culemann, W., Sasinka, C., Zezula, P., & Cenek, J. (2024). ETDD70: Eye-tracking Dyslexia Dataset [Data set]. Zenodo. https://doi.org/10.5281/zenodo.13332134

CITE THE ASSOCIATED PAPER

Sedmidubsky, J., Dostalova, N., Svaricek, R., & Culemann, W. (2024). ETDD70: Eye-tracking dataset for classification of dyslexia using AI-based methods. In Proceedings of the 17th International Conference on Similarity Search and Applications (SISAP) (pp. 1-14). Springer.

ETDD70数据集收录了70名捷克参与者的眼动追踪（eye-tracking）记录，参与者均分为诵读困难（dyslexia）者与非诵读困难读者，所有参与者年龄介于9至10岁。该数据集记录了参与者完成三项捷克语文本阅读任务时的眼动轨迹：音节阅读（任务1）、有意义文本阅读（任务4）以及伪文本阅读（任务5）。

本数据集源于本研究团队开展的“基于眼动追踪与人工智能的阅读障碍诊断”项目。该项目旨在借助人工智能工具与先进技术设备（眼动追踪），更高效地诊断阅读障碍——这是最常见的特定学习障碍之一，进而显著改善阅读障碍学生的再教育策略。本项目的核心目标是开发模型，基于上述任务中记录的眼动模式，精准区分阅读障碍者与非阅读障碍者。

数据收集工作于2022年10月至2023年8月开展，严格遵循伦理规范，已获得捷克共和国布尔诺马萨里克大学研究伦理委员会批准。

若有任何疑问或反馈，请联系我们：[e-mail] 或 [e-mail]。

ETDD70数据集可免费用于研究用途。恳请使用者在开展研究使用前联系我们（邮箱地址同上），以便我们追踪申请情况并提供相关指导。

### 参与者
本次眼动数据来自70名参与者：35名阅读障碍者与35名非阅读障碍读者，均为9至10岁的小学生（即小学四年级）。招募工作与心理辅导中心合作，协助招募已确诊阅读障碍的学生；非阅读障碍读者（无阅读障碍症状）则通过合作小学的辅导机构招募。所有参与者的法定监护人已充分知晓研究流程并同意参与研究，且获得了相应报酬。

### 刺激材料
我们基于捷克国内标准化纸质阅读障碍诊断工具设计了三项语言任务。原始文本按照眼动追踪测量需求，被转换为受控格式的数字版本（例如文本总量、字体大小、行间距、背景色等参数。

任务1：音节阅读（Syllables）包含90个音节，排列为9×10矩阵。这些音节均为捷克语常见音节。按照音节构成（基于语言学特征）分为五类：无意义开音节（辅音+元音，如“ta”“na”）、带意义的闭音节（辅音+元音+辅音，如“suk”“mák”）、由两个辅音组成的无意义音节（如“vl”“bz”）、由两个辅音结尾且带元音的无意义音节集群（如“tle”“mra”），以及由三个辅音集群构成、第三个位置带元音的有意义音节（如“mrak”“vlak”）。所有音节以黑色Times New Roman字体，显示在灰色背景上。本任务要求参与者从左至右、从上至下大声朗读所有音节。右下角设置注视十字，用于眼动触发式任务终止：当参与者注视该十字时，记录自动终止。

任务4：有意义文本阅读（MeaningfulText）是一篇关于小男孩从窗口观察松鼠的短文，面向小学三、四年级学生阅读。文本共7行，包含6个逻辑句。文本同样采用黑色字体，灰色背景、双倍行间距，右下角同样设置注视十字。本任务要求参与者完整大声朗读全文。

任务5：伪文本阅读（PseudoText）包含虚构无意义词汇，共7行15个人工语句。文本格式与注视十字设置均与有意义文本阅读任务一致。本任务要求参与者尽可能流畅地大声朗读全文。

### 眼动追踪特征
每项任务记录的原始眼动数据经进一步处理，提取基于事件的特征——注视（fixation）、眼跳（saccade），以及数十种衍生统计特征。注视点采用i2mc算法（Hessels等人，2017）检测，该算法专为儿童眼动测量的抗噪需求设计。最小注视时长设置为40ms。衍生特征提供参与者与文本交互的额外信息，分为全任务特征与感兴趣区域（Region of Interest, ROI）特征。全任务特征描述全局屏幕层面的语义信息，ROI特征则刻画小型矩形区域的局部语义特征。

各任务的特征如下：
1. 音节阅读任务
首次注视时长、平均注视时长、注视次数、不计入进出眼跳的注视与眼跳次数、回访次数——即从外部进入该感兴趣区域的传入眼跳次数。
2. 有意义文本阅读、伪文本阅读任务
全任务特征（从整个试次提取的特征：回视次数、前进眼跳与回视眼跳的比例、平均眼跳幅度、总阅读时长、平均注视时长、注视次数。
ROI特征（从分离的感兴趣区域，即行与词，提取的特征：平均注视时长、注视次数、回访次数——从外部进入该感兴趣区域的传入眼跳次数、首次注视的着陆位置。

### 人工智能分类方法
用于阅读障碍分类的基于人工智能的方法可访问：https://gitlab.fi.muni.cz/xsedmid/dyslex

### 数据集引用
Dostalova, N., Svaricek, R., Sedmidubsky, J., Culemann, W., Sasinka, C., Zezula, P., & Cenek, J. (2024). ETDD70: Eye-tracking Dyslexia Dataset [数据集]. Zenodo. https://doi.org/10.5281/zenodo.13332134

### 关联论文引用
Sedmidubsky, J., Dostalova, N., Svaricek, R., & Culemann, W. (2024). ETDD70: 基于人工智能方法的阅读障碍分类眼动数据集。收录于第17届相似性搜索与应用国际会议（SISAP）会议论文集（第1-14页）。Springer出版社。