AI-based identification of amblyopia from fixation eye movement digital biomarkers

Amblyopia is a neurodevelopmental disorder caused by decorrelated binocular input early in life and is characterized by reduced visual acuity, impaired stereopsis, and decreased fixation stability. Clinical diagnosis still relies largely on subjective visual-acuity assessment, which is particularly challenging in young children and in patients with co-occurring nystagmus, contributing to misclassification and delayed treatment. Although prior work has quantified fixation eye-movement abnormalities in amblyopia, these approaches have not consistently achieved reliable, automated separation of patients and controls. Here we present an artificial-intelligence framework for objective classification of amblyopic patients using quantitative fixation-eye-movement features. We analyzed a high-dimensional dataset comprising 510 features extracted from fixation recordings of 123 participants (86 amblyopic, 35 controls), split into training (70%) and testing (30%) sets. A custom genetic-algorithm (GA) feature-selection method, using a two-hidden-layer neural network as the fitness function and explicitly favoring solutions with fewer features, identified 9 discriminative features spanning fast and slow eye-movement metrics. Using these features, the classifier achieved 100% accuracy on both training and test sets. Importantly, 5-fold cross-validation (10 repetitions, 500 independent network trainings) on the 86-subject training set yielded accuracy 85.8 ± 11.4%, sensitivity 74.8 ± 23.7%, specificity 94.1 ± 8.6%, precision 87.5 ± 21.9%, weighted F1 = 0.847 ± 0.137, supporting the model’s robustness in small-sample, high-dimensional settings. Unlike deep-learning methods that typically require large datasets, this hybrid computational-intelligence approach efficiently explores complex feature spaces and offers a promising framework for objective, data-driven diagnostic support in amblyopia.