Abstract
The increasing burden of childhood myopia creates a need for screening models that remain informative when routinely collected clinical measurements are incomplete. We propose SSR-Stacking, a hybrid framework integrating Graph Neural Networks (GNN) with Ensemble Learning for classification of final-follow-up high-myopia status under simulated missingness of baseline spherical equivalent (SE). Unlike traditional models, we construct a Social-Environment Graph based on school-class affiliations to represent classroom-level relational structure. Our framework features a Self-Supervised Reconstruction (SSR) mechanism toreconstruct randomly masked baseline SE from neighboring peers and a Meta-Stacking layer to fuse GNN embeddings with diverse ML classifiers (XGBoost, SVM, RF). Evaluated on a longitudinal dataset (
N
= 4, 973) from Binchuan, China, under 50% simulated baseline-SE missingness using classroom-level grouped folds and out-of-fold stacking, SSR-Stacking achieved a ROC-AUC of 0.9184 (95% CI: 0.9062–0.9306), PR-AUC of 0.6120, Recall of 0.8119, Specificity of 0.9851, and F1-score of 0.7506 on unsampled held-out predictions with a natural high-myopia prevalence of 4.06%. These internal results support further evaluation under natural missingness and external validation before clinical deployment.