Development of a deep learning model to predict cycloplegic myopia and pre- myopia: based on a large-scale screening of school-aged children.

purposeThis study aims to develop and validate practical deep learning (DL) and machine learning (ML) models to predict cycloplegic myopia, pre-myopia, and non-myopia in school-aged children using non-cycloplegic parameters, thereby facilitating early detection and intervention.

methodsA cross-sectional study enrolled 3,077 children aged 6-10 years from Tianjin, China. Ocular parameters, including non-cycloplegic dioptre sphere (DS), axial length (AL), corneal curvature (K1, K2), and sex, were incorporated into four single multi-class classification models: CatBoost, XGBoost, Random Forest (RF), and TabPFN (a tabular DL model). Performance was evaluated using AUC, accuracy, sensitivity, specificity, and SHAP analysis for interpretability.

resultsThe TabPFN model demonstrated slightly higher overall performance (AUC = 0.89), with AUCs of 0.93 for myopia and 0.78 for pre-myopia prediction. AL emerged as the most influential feature, while K2 contributed to model predictions alongside other biometric parameters. Decision curve analysis demonstrated clinical utility for combined pre-myopia+myopia classification at intermediate thresholds.

conclusionThe TabPFN model outperformed traditional ML models in accurately classifying myopia and pre-myopia using non-cycloplegic parameters, offering a practical tool for large-scale school screenings. AL was the dominant contributor to model predictions, while K2 provided complementary information within the multivariable framework.