To develop and validate an Expert-Guided Stacking (EGS) predictive framework for early risk stratification of high-myopia using longitudinal school screening data.
Approach:
Model Development: Constructed predictors from screening records preceding outcome-defining follow-up records, evaluated performance across six classifiers, and proposed a hybrid EGS model integrating multi-model ensemble architecture with a clinical risk-heuristic override.
Validation: Used student-level partitioning, 5-fold cross-validation for tuning, and held-out test-set evaluation for performance assessment.
Key Findings:
AdaBoost achieved high overall discrimination (AUC = 0.9992).
EGS framework demonstrated high Recall (0.9533) and Precision (0.9211).
SHAP analysis confirmed the importance of longitudinal trajectory features for model interpretability.
Interpretation:
The EGS framework provides a robust solution for school-based myopia surveillance, focusing on high-risk recall.
Limitations:
The study is based on data from a specific region (Binchuan County, China), which may limit generalizability.
The reliance on non-cycloplegic measures may introduce misclassification in younger children.
Conclusion:
EGS-Net enhances predictive sensitivity for high-myopia risk by leveraging longitudinal data and clinical heuristics.
