Skip to Content
DiagnosticsDiagnostics
  • Article
  • Open Access

14 November 2026

29 Pages

Word-Anchored Eye-Movement Measures for Interpretable and Calibration-Robust Dyslexia Screening on the ETDD70 Dataset

and
1
Department of Computer Science, College of Computer and Information Sciences, Jouf University, Sakaka 72388, Saudi Arabia
2
Department of Software Engineering, College of Computer and Information Sciences, Jouf University, Sakaka 72388, Saudi Arabia
*
Author to whom correspondence should be addressed.

Abstract

Background/Objectives:Eye-tracking models for dyslexia screening usually describe gaze in screen coordinates, although dyslexia is a difficulty with words and letters. We aimed to develop an interpretable word-anchored screening approach that is robust to calibration error. Methods: We analysed the public ETDD70 dataset (70 Czech children, 35 with dyslexia). The position of all 292 stimulus words was recovered exactly from the released landing positions, and every fixation was assigned to a word and letter. Each child was described by word-level reading measures, linguistic sensitivities and saccade geometry in letters (LexiGaze), complemented by empirical-Bayes profiles and a generative likelihood-ratio classifier. Models were evaluated by nested participant-level cross-validation and under simulated gaze error. Results: The resulting word assignment matched the dataset labels for 99.99% of fixations. Dyslexia was associated with 42% longer gaze durations on meaningful text and with an exaggerated word-length effect on fixation counts. LexiGaze reached a balanced accuracy of 82.57% (AUC 0.913) and the generative classifier an AUC of 0.938, comparable to existing descriptors. Under gaze noise of 30 pixels, pixel-based geometry lost 23.4 points of balanced accuracy and LexiGaze 6.0 (direct comparison p = 0.053); with the recorded vertical coordinates, without drift correction, they lost 5.1 and 0.3 points. Conclusions: Word-anchored measures match existing accuracy while being interpretable in psycholinguistic terms and less sensitive to calibration error.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.