Optimizing Myopia Prediction in Children and Adolescents Using Machine Learning: A Multi-Factorial Risk Assessment Model

Yue Xi¹Wei Zhu²Wenjing Yan³Xiaobin Wei⁴Qian Zhao⁵Shenting Dai^6*

• ¹School of Physical Education and Health, East China Normal University, Shanghai, China
• ²Yinchuan Maternal and Child Health Hospital, Yinchuan, China
• ³Shanxi Normal University, Taiyuan, China
• ⁴Tsinghua University, Beijing, China
• ⁵The No. 10 Primary School of Zhongning County, Zhongwei, China
• ⁶East China University of Science and Technology, Shanghai, China

Background Most previous studies on myopia in children and adolescents have primarily focused on genetic and environmental factors. This study aimed to explore modifiable behavioral, sociodemographic, and psychological contributors to myopia and to evaluate the potential of machine learning (ML) models in identifying at-risk individuals. Methods A cross-sectional survey was conducted in eight primary and secondary schools in a Chinese province between October and December 2023. The dataset was split into training and validation sets (7:3). LASSO regression identified potential predictors, followed by multivariate logistic regression to determine independent risk factors. Nine machine learning algorithms were used to build prediction models: logistic regression, support vector machine (SVM), gradient boosting machine (GBM), neural network (NNET), extreme gradient boosting (XGBoost), k-nearest neighbors (KNN), adaptive boosting (AdaBoost), LightGBM, and CatBoost. Model performance was evaluated using accuracy, F1 score, specificity, sensitivity, and area under the receiver operating characteristic (ROC) curve (AUC). SHapley Additive exPlanations (SHAP) were used to interpret variable contributions in the best-performing model. Results The study included 2,086 children and adolescents (mean age 9.8 ± 2.7 years; 50.5% female), with an overall myopia prevalence of 25.12%. Independent risk factors for myopia included parental myopia, only-child status, physical activity level, mother’s education level, age, and physical activity behavior. Among all models, the LightGBM algorithm achieved the best predictive performance (AUC = 0.738, 95% CI: 0.709-0.767). SHAP analysis identified parental myopia, physical activity level, only-child status, and physical activity behavior as the most influential predictors. Conclusion Although ML models showed limited predictive accuracy, they helped identify modifiable risk factors associated with childhood and adolescent myopia. These findings may inform the design of targeted prevention strategies and early behavioral interventions rather than serve as clinical diagnostic tools.