Integrating Multi-Task Eye Tracking and Interpretable Machine Learning for High-Accuracy Screening of Amblyopia in Pediatric Populations

Amblyopia is a developmental disorder of spatial vision in which abnormal visual experience leads to persistent reductions in acuity and contrast sensitivity, even after optimal optical correction. We introduce a brief, child-friendly battery of task-evoked eye tracking that probes fixation stability, fine pattern processing, and smooth pursuit control across three simple paradigms. Oculomotor traces are transformed into physiologically interpretable markers-fixation dispersion and saccadic strategy, orientation-dependent drift and stability, pursuit gain, and tracking error-and used to train a compact classifier with subject-wise validation and probability calibration. In a cohort of school-aged participants with clinically diagnosed unilateral amblyopia and age-matched visually normal controls tested under best-corrected viewing conditions, the approach consistently separated groups with stable performance across folds; feature-importance analyses indicated that pursuit- and orientation-dependent markers contributed most. The protocol runs in minutes, is objective and noninvasive, and is well tolerated in pediatric settings. By quantifying functional consequences of amblyopic vision that complement conventional acuity testing, this work positions task-evoked eye movements as practical biomarkers for screening and monitoring, and lays the groundwork for prospective validation and age-stratified norms in community and school-based vision care.