Explainable and Uncertainty-Aware Eye-Tracking for Dyslexia-Related Reading Support

Abstract

Early identification of dyslexia-related reading difficulties is essential for timely educational support, but automated approaches must remain interpretable, uncertainty-aware and subordinate to professional assessment. This study presents an explainable eye-tracking framework for analysing dyslexia-related reading patterns across syllable, meaningful-text and pseudo-text tasks. A publicly available dataset comprising 70 children, including 35 dyslexic and 35 non-dyslexic readers, was evaluated using participant-independent five-fold cross-validation. Fixation, saccade, pupil and reading-duration features were extracted separately from each task and eye-movement event family. Logistic Regression, Support Vector Machine, Random Forest and Extra Trees classifiers were compared with EduGaze-Select, a task-aware ensemble that combined task-specific predictions using validation-derived fusion weights. Split conformal prediction was incorporated to identify uncertain cases requiring further educational assessment. Logistic Regression achieved the strongest overall performance, with 90.00% accuracy, 90.00% balanced accuracy, an ROC-AUC of 0.9347 and a PR-AUC of 0.9534. EduGaze-Select achieved 82.86% balanced accuracy and an ROC-AUC of 0.9061 while providing task-level interpretability and calibrated referral support. The conformal component attained 94.29% empirical coverage against a nominal target of 90%, produced unique decisions for 74.29% of participants and referred 25.71% for further assessment. These findings demonstrate that explainable and uncertainty-aware eye-tracking analysis can support dyslexia-related reading assessment while preserving human oversight and avoiding automated diagnostic claims.