TC-GAFnet: A Time-Aware Contrastive Grouped Attention Fusion Network with Empirical Priors for Adolescent Vision Prediction

Abstract

Accurate prediction of spherical equivalent (SE) values is essential for early intervention in adolescent myopia management. Large-scale vision screening data provide valuable opportunities for longitudinal prediction; however, their utilization is challenged by sparse observations, irregular sampling intervals, and limited sequence lengths. To address these challenges, this study proposes TC-GAFnet, a medical prior-augmented temporal representation learning framework for robust SE prediction from irregular longitudinal screening data. TC-GAFnet integrates multi-dimensional representation learning, dual-stream attention, temporal encoding, and multi-view contrastive alignment to capture physiological progression patterns and enhance prediction robustness. Specifically, temporal encoding mitigates the influence of non-uniform sampling intervals, while multi-view contrastive learning achieves semantic alignment between binocular physiological trajectories. Experiments on 75,952 records demonstrate that TC-GAFnet achieves an MAE of 0.4882 D and an R2 of 0.8774, with an 82.63% sensitivity for high-myopia screening. SHAP analysis further indicates that learned feature representations are consistent with clinical priors, supporting the model’s interpretability and potential for reliable SE prediction.