Revolutionizing Pediatric Myopia Care: A Machine Learning Approach for Rapid and Accurate Pre-clinical Screening

Background/Objective: Myopia has become a prominent public health issue in China, significantly impacting the visual health of children and adolescents. The condition is characterized by a high incidence rate, increasing prevalence, and a trend toward earlier onset, highlighting the critical need for early and accurate diagnosis. Current clinical diagnostic methods primarily depend on subjective evaluations by optometrists and the use of isolated parameters, leading to inefficiencies and inconsistent outcomes. Moreover, there remains a lack of diagnostic tools that can effectively integrate multi-parameter analysis while ensuring robust data privacy protection. This study aims to develop an artificial intelligence (AI) diagnostic model that achieves objective, accurate, and safe diagnosis of myopia in children without cycloplegia through multi-parameter fusion and to enable local deployment. The proposed model is intended to be a reliable tool for clinical applications and large-scale screening projects, while ensuring strong protection of patient privacy. Methods: We built a transparent, rule-driven AI framework using clinical guidelines. Key ocular parameters—visual acuity, spherical equivalent, axial length, corneal curvature, and axial ratio—were encoded as logical rules in Python and incorporated via instruction fine-tuning. The model was trained and validated on retrospective clinical data (70% training, 15% validation, 15% test) using five algorithms: gradient boosting, logistic regression, random forest, SVM, and XGBoost. Performance was evaluated using accuracy, precision, recall, F1 score, and mean AUC across classes. Results: The model classifies refractive status into five categories: hyperopia, pre-myopia, mild, moderate, and high myopia. All five different algorithms demonstrated excellent diagnostic and classification performance. Gradient boosting achieved the best overall performance, with an accuracy of 98.67%, an F1 score of 98.67%, and a mean AUC of 0.957—outperforming all other models. Conclusions: This study successfully developed an artificial intelligence-based myopia diagnosis system for children under non-dilated pupil conditions. The system is interpretable and privacy-preserving, and has excellent diagnostic and classification performance, making it suitable for clinical decision support and large-scale screening applications. It has great potential to promote the development of early intervention, precision prevention, and control strategies for childhood myopia.