Interpretable machine learning for self-reported reduced uncorrected visual acuity in school students

Myopia and reduced vision among children and adolescents place a substantial burden on public health, particularly where eye-care resources are limited. We evaluated an interpretable questionnaire-based framework for modelling self-reported reduced uncorrected visual acuity.

The analysis included the first eligible record for each of 7,118 students from seven schools across three survey rounds. Sixteen demographic, familial and behavioural predictors were used to model an affirmative response to a question about uncorrected visual acuity below 5.0 in at least one eye. Internal-external cross-validation (IECV) comprised seven outer folds, each holding out one school. Preprocessing, model selection, probability calibration and Youden threshold estimation were confined to the training schools.

The selected workflow achieved an area under the receiver operating characteristic curve (AUC) of 0.706 [95% confidence interval (CI) 0.694-0.718], a Brier score of 0.212, sensitivity of 0.730 and specificity of 0.575. The academic-year-only benchmark achieved an AUC of 0.611 and a Brier score of 0.233. Stage-specific AUCs ranged from 0.633 to 0.680, with substantial variation in sensitivity and specificity. Applying the same development procedure to the full cohort selected CatBoost. SHapley Additive exPlanations (SHAP) identified academic year, parental myopia and desk-chair adjustment as leading contributors to its predictions.

Questionnaire responses provided predictive information about self-reported visual status beyond academic year, with moderate overall discrimination and variation across educational stages. These findings provide a basis for further evaluation of the framework in school eye-health research.