Abstract
Background/Objectives: Efficient visual processing is relevant for reading, writing, and sustained attention in schoolchildren. However, the relative discriminative value of different visual domains for identifying teacher-reported academic difficulties remains unclear. This study evaluated five visual system alteration domains for identifying teacher-reported academic difficulties using machine learning models. Methods: An observational analytical study was conducted in 581 primary schoolchildren. After complete-case analysis, 506 participants were included in the machine-learning analyses. Academic functioning was rated by teachers using a 1–5 ordinal scale and dichotomized as a pragmatic school-based indicator of teacher-reported academic difficulties. Five predictor groups were analyzed: DIVE-based oculomotor function, clinical oculomotor assessment, accommodative system, vergence system, and axial length. Five classifiers were evaluated using stratified 5-fold cross-validation combined with model-complexity penalization through hyperparameter optimization. Results: The accommodative system showed the highest cross-validated performance (XGBoost: accuracy 0.952 ± 0.021; macro-F1 0.919 ± 0.028), followed by clinically assessed and instrumentally assessed oculomotor predictors. Oculomotor alterations also showed strong performance, whereas vergence alterations showed high specificity but very low sensitivity, and abnormal axial length showed limited discriminative performance. Conclusions: Functional visual domains, particularly accommodation and oculomotor control, showed stronger cross-validated classification performance than vergence or axial-length variables. These findings are exploratory and require external validation cohorts, standardized academic outcomes, and future combined-domain modeling before clinical or educational implementation.