Answers are grounded in this paper and cite their sources.
Background Rising global myopia rates are a major public health challenge because of the risk of irreversible vision loss in high myopia, and this creates a need for early identification of children on rapidly progressing trajectories. School-based screening characterises a child’s current refractive state but does not forecast where that trajectory is heading. We developed and internally validated a time-aware deep learning model that forecasts refractive progression directly from irregularly sampled school screening records, and assessed whether school-level geospatial context adds prognostic information. Methods A Time-Aware Transformer was adapted to process irregularly sampled refractive screening sequences without imputation. It combines an Adaptive-Frequency Time Embedding, which tunes a sinusoidal frequency spectrum to clinical sampling intervals using two learnable parameters, with a temperature-scaled Gated Residual Network that selectively integrates school-level environmental indices. The model was developed on a Gold Standard cohort of 2874 students from Yecheng County, China with complete four-year screening trajectories (2020 to 2023). Outcomes were one-year-ahead spherical equivalent refraction (SER) and classification of rapid myopic progression ( \:Δ\: SER \:≤\:-0.75 D/year). Internal validation used a held-out test set, five-fold cross-validation, a ten-seed stability analysis, a school-level grouped split and leave-one-school-out validation. Discrimination, calibration and net benefit were assessed, and nine baseline models were compared using paired test-set bootstrap tests with post hoc multiplicity control. Results The model achieved a mean squared error (MSE) of 0.168 D 2 and an \:R² of 0.929 for one-year-ahead SER, and an area under the receiver operating characteristic curve of 0.891 for identifying rapid progressors, with good calibration (calibration slope 0.968, intercept $$\:-0.014$$ D) and higher net benefit than referring all or no children across referral thresholds of 0.10 to 0.55. Performance was stable in five-fold cross-validation (MSE 0.171 ± 0.017) and across ten random seeds (MSE 0.168 ± 0.004). Accuracy was comparable to the strongest baselines rather than clearly superior: the difference from ContiFormer was 0.010 D in root mean squared error, well below the precision of non-cycloplegic autorefraction, while the proposed model used 42% fewer parameters (52 K versus 89 K) and trained 2.7 times faster. When whole schools were held out, MSE rose to 0.191 and the benefit attributable to environmental features fell from 0.027 to 0.005 D 2 , indicating that most of the apparent geospatial signal reflects school identity rather than measured environment. Prediction was hardest for children who converted to myopia during follow-up (MSE 0.243) and easiest for persistent non-myopes (MSE 0.105). Conclusion A compact time-aware transformer forecasts one-year-ahead refractive error from routinely collected school screening data with accuracy comparable to substantially larger continuous-time models, with calibrated risk estimates and clinically interpretable attention. The case for this architecture rests on efficiency, stability and interpretability rather than on an accuracy advantage. School-level environmental indices should be understood as geospatial baseline priors rather than as individual-level exposures.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: