Machine Learning Prediction of Binocular Vision Recovery in Accommodative Esotropia: A Prospective Multicenter Study

Abstract

Introduction: Accommodative esotropia is the most common childhood convergent strabismus, yet predicting binocular vision recovery after treatment remains challenging. This study developed and validated machine learning (ML) models to predict binocular vision recovery using baseline clinical parameters.

Methods: This prospective multicenter study enrolled 156 patients (aged 2–12 years) with accommodative esotropia across three private hospitals in Indonesia. The unit of analysis was patients. Baseline binocular vision parameters were used to train four ML models (gradient boosting, random forest, neural network, logistic regression) with 5-fold stratified cross-validation. Treatment success was defined as stereoacuity ≤100 arc seconds at 12 months. Model performance was evaluated using AUC-ROC and SHAP feature importance.

Results: Treatment success was achieved in 104 patients (66.7%). Gradient boosting achieved the highest AUC of 0.903 (95% CI: 0.854–0.952; sensitivity 0.875; specificity 0.827). The strongest predictors were baseline stereoacuity ≤400 arc seconds (OR = 4.15; p < 0.001), deviation angle ≤20 PD (OR = 3.42; p < 0.001), and Worth 4-dot fusion (OR = 3.21; p = 0.001).

Conclusion: ML models accurately predicted binocular vision recovery in accommodative esotropia, identifying clinically interpretable predictors that may optimize treatment selection in pediatric strabismus management.