JMIR Publications

ABSTRACT

Background:

Myopia has escalated into a critical public health crisis among children and adolescents in China. While numerous studies have explored risk factors using traditional statistical methods, there remains a challenge in handling high-dimensional behavioral data and accurately identifying the most predictive variables for precision prevention.

Objective:

This study aimed to investigate the prevalence of myopia and identify key influencing factors among primary and secondary school students in Shangrao, China, using machine learning (ML) models.

Methods:

A school-based cross-sectional study was conducted in October 2024, involving 22,359 students from grades 4 to 12. Data on demographics, visual acuity, non-cycloplegic autorefraction, and behavioral factors were collected. A multi-stage feature selection process, integrating univariate logistic regression with four ML algorithms (LASSO, RF, XGBoost, and LightGBM), was employed to identify the most predictive variables. The optimal logistic regression model was used to construct a clinical nomogram. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis (DCA). Finally, model interpretability was enhanced using Shapley Additive Explanations (SHAP) to quantify the impact of each feature.

Results:

The overall prevalence of myopia was 58.51%. Prevalence increased significantly with grade level, ranging from 38.15% in upper primary school to 77.74% in senior high school. Females (63.40%) and students with a parental history of myopia (62.07% for one parent, 68.32% for both) exhibited a significantly higher prevalence (all P < 0.001). Among the ML models, the logistic regression model demonstrated the best predictive performance (AUC = 0.726, 95% CI: 0.718–0.734) and was visualized as a nomogram incorporating nine key predictors. The nomogram showed robust discriminative ability (AUC = 0.725), good calibration, and provided net clinical benefit across a wide threshold probability range (25–90%). SHAP analysis revealed that grade level, parental history of myopia, and gender were the most critical predictors. Key modifiable protective factors included maintaining a proper reading distance and spending recess outdoors.

Conclusions:

The prevalence of myopia is high among students in Shangrao. The ML-derived nomogram serves as a practical tool for risk assessment, effectively identifying both established and modifiable risk factors. Our findings support a precision prevention strategy focusing on high-risk groups while promoting behavioral interventions, such as ensuring adequate reading distance and encouraging outdoor recess.