Abstract / Summary
Pediatric myopia screening captures refractive and biometric measurements but rarely converts them into individualised risk estimates. Whether automated biometric features (pupil diameter, interpupil distance, gaze deviation) add predictive value beyond standard refraction is unevaluated. We developed and internally validated a calibrated multi-outcome risk model using 4,220 screening records from 2,373 children (ages 6–13) from 2018 to 2021. Six outcomes were predicted: high myopia (spherical equivalent $$\:\le\:-3.00$$ D) at 1, 2 and 3 years, mild-to-moderate progression, myopia onset, and rapid progression. Progressively richer feature tiers were compared with patient-grouped cross-validation, patient-clustered bootstrap, DeLong tests with Holm-Bonferroni correction, temporal validation, six comparators, and a pre-specified sex-stratified fairness assessment. Reporting followed TRIPOD + AI. The clinical refraction model (age, sex, spherical equivalent) achieved AUC 0.878 (95% CI 0.815–0.931) for 1-year high myopia, 0.828 at 2 years and 0.795 at 3 years; mild-to-moderate progression reached 0.782 and onset 0.739. Temporal validation preserved discrimination, and no comparator was significantly superior. Biometric features added no discrimination on any outcome, no single feature changing AUC beyond 0.005. After isotonic recalibration, decision curve analysis showed net benefit; at a 30% threshold the 2-year model reached 83.4% sensitivity and 97.1% negative predictive value, without disparity by sex. A parsimonious logistic regression model using only routine screening data predicts pediatric high myopia across multiple horizons. Screening-device biometric features add no incremental value, being correlated with the dominant age and refractive predictors.