EGS-Net: a knowledge-augmented machine learning framework for predicting future high-myopia risk from longitudinal school-screening trajectories.

The rapid increase in childhood myopia highlights the need for accurate, non-invasive tools for early risk stratification in large-scale screening; however, static cross-sectional data often fail to capture dynamic refractive trajectories. In this study, we developed and validated an Expert-Guided Stacking (EGS) predictive framework using longitudinal school screening data (Autumn 2023-Autumn 2025) from Binchuan County, China. For each student, predictors were constructed only from screening records preceding the outcome-defining follow-up record, thereby preserving the original temporal prediction boundary. We first evaluated performance across six conventional classifiers (LR, RF, XGBoost, SVM, NB, and AdaBoost), then proposed a hybrid EGS model that integrates a multi-model ensemble architecture with a clinical risk-heuristic override module. This framework was specifically designed to reduce false negatives and improve prediction of future high-myopia risk by leveraging historical longitudinal refractive trajectories. Model development used student-level partitioning, with 5-fold cross-validation for tuning and held-out test-set evaluation for final performance assessment. Although AdaBoost attained a high overall discrimination (AUC = 0.9992), the proposed EGS framework achieved clinically favorable utility with high Recall (0.9533) and Precision (0.9211), enabling reliable future-risk identification while avoiding the false-positive burden of less precise high-recall models. SHAP analysis verified the critical contribution of longitudinal trajectory features to model interpretability and transparency. Our findings demonstrate that this knowledge-augmented ML approach delivers a robust, scalable solution for school-based myopia surveillance, with a priority on high-risk recall to support timely clinical intervention and personalized vision care.