Comprehensive behavioral and demographic correlates in myopia risk among Tibetan adolescents: a machine learning...

Myopia is a global public health concern, especially among children and adolescents. This exploratory study aimed to develop a myopia risk prediction model for Tibetan adolescents using interpretable machine learning, while explicitly acknowledging the methodological limitations of non-cycloplegic refraction and self-reported data. We included 735 Tibetan adolescents. An XGBoost algorithm was used to build a binary classification model, with the SHAP method applied for feature interpretation. Data were split into 80% training and 20% test sets to analyze associations between behavioral, demographic factors, and myopia risk. Myopia was defined using non-cycloplegic autorefraction (uncorrected visual acuity < 5.0 [logMAR 0.0] and spherical equivalent ≤ - 0.50 D). In children aged 9-15 years, non-cycloplegic autorefraction is known to overestimate myopia due to accommodative tone, leading to outcome misclassification. Moreover, behavioral exposures were self-reported by children, introducing recall and desirability bias. Consequently, all findings are purely hypothesis-generating and require validation using cycloplegic refraction and objective exposure measures. Statistical analyses included independent samples t-tests, Mann-Whitney Utests, chi-square tests, and binary logistic regression. Older age was significantly associated with higher myopia risk (OR = 1.200, 95% CI: 1.069-1.347, P = 0.002). Male sex (OR = 0.583, 95% CI:0.427-0.796) and boarding at school (OR = 0.419, 95% CI: 0.250-0.702) were associated with lower myopia prevalence. Behavioral associated factors included irregular breakfast habits (OR = 2.882), reduced frequency of classroom seat rotation, and longer daily after-school study time. Machine learning analysis identified learning behaviors (seat rotation, study time, grade) as the most important predictors, followed by lifestyle factors (breakfast habits, screen time). However, due to the absence of cycloplegia and the cross-sectional design, these associations may be biased and should not be interpreted causally. In this exploratory cross-sectional study with major methodological limitations, several behavioral and demographic factors were associated with myopia status among Tibetan adolescents. The findings do not establish causality and are not suitable for clinical or public health recommendations. Future research must employ cycloplegic refraction, objective exposure monitoring (e.g., wearable light sensors, school records), and longitudinal designs before any reliable conclusions can be drawn.