Machine Learning-Based Prediction of Strabismus with and without Amblyopia from Fixation Eye Movements

Area of Science:

• Ophthalmology
• Computer Science
• Machine Learning

Background:

• Strabismic amblyopia and strabismus are common visual impairments.
• Objective diagnostic methods are crucial for timely intervention.
• Current diagnostic approaches may have limitations in differentiating these conditions.

Purpose of the Study:

• To evaluate machine learning (ML) models trained on fixation eye movement (FEM) features.
• To determine if ML models can differentiate strabismic amblyopia, strabismus without amblyopia, and visually normal controls.
• To identify key FEM features predictive of these conditions.

Main Methods:

• A cross-sectional study involving 159 subjects (amblyopia, strabismus, controls).
• Recorded horizontal and vertical eye-position traces during monocular viewing.
• Computed various FEM metrics including fixation instability, eye deviation, and vergence instability.
• Trained multiple ML models for six binary classification tasks.
• Utilized Shapley Additive exPlanations (SHAP) to identify predictive features.

Main Results:

• ML models achieved high accuracy in distinguishing moderate/severe strabismic amblyopia from strabismus (0.81 ± 0.09) and controls (0.90 ± 0.07).
• Strong performance was also observed in classifying strabismus versus controls (0.90 ± 0.07) and amblyopia versus controls (0.90 ± 0.07).
• SHAP analysis indicated that viewing amblyopic eye abnormalities were key predictors of amblyopia, while nonviewing-eye FEM metrics and binocular coordination features predicted strabismus.

Conclusions:

• FEM features combined with ML can objectively distinguish strabismic amblyopia from strabismus and controls.
• The strongest performance was noted in moderate/severe disease.
• Future research should focus on younger populations and portable eye-tracking for scalable screening.