Deploying Machine Learning Strategies on Smartphones for Simplified Myopia Screening among School-Aged Children (Preprint)

BACKGROUND Myopia is a growing global public health concern, with particularly high prevalence among school-aged children in East and Southeast Asia and increasing risk of sight-threatening complications in high myopia. Early identification of premyopia is critical for timely intervention, yet current screening methods rely on specialized equipment or static imaging and fail to capture dynamic near-work behaviors, limiting accessibility and scalability. Therefore, an accessible and behavior-aware screening approach is urgently needed. OBJECTIVE To validate a smartphone-based machine learning (ML) method for home myopia screening in school-aged children, focusing on translational utility in resource-limited settings and premyopia detection, addressing gaps in static tools. METHODS A total of 150 school-aged children (6–18 years) were enrolled for ML model training/validation, with 54 additional eyes for preliminary external testing. Sample size was justified via power analysis. Smartphone-acquired features included age, sex, pupil distance, eye-screen distance, and cohesion angle. Pixel-to-distance calibration and measurement repeatability were validated. Stratified tenfold repeated cross-validation and bootstrapping assessed model stability. ML models predicted spherical equivalent (SE) and classified myopia (SE≤-0.50 D) vs. premyopia (SE: -0.50 D to +0.75 D); SHAP quantified feature importance. RESULTS Participants (mean age 9.24 ± 2.23 years) had a 61.3% myopia rate. Eye-screen distance was the top feature (importance=1.00). Random forest performed best: SE prediction (test set: R²=0.523, 95% CI 0.237–0.802; MAE=0.686 D, 95% CI 0.480–0.890) and myopia classification (test set: AUC=0.855, 95% CI 0.716–0.976; accuracy=0.779). Bootstrapped CV