[Paper Review] DysLexML: Screening Tool for Dyslexia Using Machine Learning

Thomais Asvestopoulou, Victoria Manousaki+5 more|arXiv (Cornell University)|Mar 14, 2019

Reading and Literacy Development8 references19 citations

TL;DR

DysLexML is a machine learning-based screening tool for developmental dyslexia that analyzes eye-tracking data from children reading text, using a minimal set of four key features—mean and median saccade length, number of short forward movements, and number of multiply fixated words—achieved via LASSO-based feature selection. It achieves 97% accuracy with a linear SVM classifier and demonstrates robust performance under artificial fixation noise, enabling deployment in low-cost, real-world settings with consumer-grade eye-trackers.

ABSTRACT

Eye movements during text reading can provide insights about reading disorders. Via eye-trackers, we can measure when, where and how eyes move with relation to the words they read. Machine Learning (ML) algorithms can decode this information and provide differential analysis. This work developed DysLexML, a screening tool for developmental dyslexia that applies various ML algorithms to analyze fixation points recorded via eye-tracking during silent reading of children. It comparatively evaluated its performance using measurements collected in a systematic field study with 69 native Greek speakers, children, 32 of which were diagnosed as dyslexic by the official governmental agency for diagnosing learning and reading difficulties in Greece. We examined a large set of features based on statistical properties of fixations and saccadic movements and identified the ones with prominent predictive power, performing dimensionality reduction. Specifically, DysLexML achieves its best performance using linear SVM, with an a accuracy of 97 %, with a small feature set, namely saccade length, number of short forward movements, and number of multiply fixated words. Furthermore, we analyzed the impact of noise on the fixation positions and showed that DysLexML is accurate and robust in the presence of noise. These encouraging results set the basis for developing screening tools in less controlled, larger-scale environments, with inexpensive eye-trackers, potentially reaching a larger population for early intervention.

Motivation & Objective

• To develop a scalable, accurate, and robust screening tool for developmental dyslexia using machine learning and eye-tracking data.
• To identify a minimal, interpretable set of oculomotor features with high predictive power for dyslexia diagnosis.
• To evaluate the robustness of the model under realistic noise conditions common in low-cost eye-tracking devices.
• To enable large-scale, low-cost screening in less controlled environments such as schools or homes.
• To support early intervention by identifying at-risk children efficiently and non-invasively.

Proposed method

• Eye-tracking data from 69 Greek-speaking children (32 diagnosed with dyslexia) were collected during silent reading of two texts using a custom eye-tracker at 60Hz.
• Fixations were detected using a dispersion algorithm, and 148 statistical features were extracted based on fixation and saccade properties.
• LASSO regression with 1-standard error selection (λ1SE) was applied for dimensionality reduction, identifying the most predictive features.
• Multiple machine learning classifiers, including linear SVM, Naïve Bayes, and others, were trained and compared on the reduced feature set.
• Artificial Gaussian noise was added to fixation positions (σ = 1 to 40 pixels) to simulate real-world tracking errors and assess model robustness.
• Performance was evaluated using accuracy, with the best model achieving 97% on the baseline text using only four features.

Experimental results

Research questions

• RQ1 Can a machine learning model achieve high accuracy in distinguishing dyslexic from typical readers using only a small set of oculomotor features from eye-tracking data?
• RQ2 How does the performance of the model degrade under realistic fixation noise, and is it robust enough for deployment with low-cost eye-trackers?
• RQ3 Which specific eye movement features are most predictive of dyslexia, and how do they vary with text difficulty?
• RQ4 Can a minimal feature set be identified that maintains high diagnostic accuracy while improving interpretability and scalability?
• RQ5 To what extent does text difficulty influence the selection of predictive features in dyslexia screening models?

Key findings

• The linear SVM classifier achieved the highest accuracy of 97% on the baseline text using only four features: mean and median saccade length, number of short forward movements, and number of multiply fixated words.
• The model remained robust under artificial fixation noise, maintaining high accuracy for noise levels up to σ = 40 pixels, indicating suitability for low-cost eye-trackers.
• Feature selection via LASSO with λ1SE successfully reduced the feature space without sacrificing accuracy, resulting in a highly interpretable and minimal model.
• For the easier text, only two features—mean saccade length and number of short forward movements—were selected, indicating that text difficulty reduces the relevance of word-specific features.
• The study demonstrates that a minimal, highly interpretable feature set can achieve state-of-the-art performance, outperforming other studies in feature set size while maintaining comparable or better accuracy.
• The results support the feasibility of deploying such a system in real-world, large-scale screening environments using affordable, non-specialized eye-tracking hardware.