**Background:** High myopia (spherical equivalent ≤ -6 D) increases lifetime risk of complications such as cataract, retinal detachment, myopic choroidal neovascularization, and glaucoma. The global burden is projected to reach 938 million people by 2050, with East Asia facing the highest potential productivity loss. In southern China, prevalence patterns of high myopia in children and adolescents remain unclear. Early identification of high-risk individuals is critical to reduce occurrence and progression. This study aimed to determine high myopia prevalence using real-world screening data and to predict its onset via machine learning.
**Methods:** This retrospective school-based study was conducted in 13 cities of Guangdong Province, southern China, with varying gross domestic products (GDP). Data were collected from electronic vision screening (Guangdong Eyevision Medical Technology Co, Ltd) between 2019 and 2021. Inclusion criteria: children and adolescents aged 6-20 years with complete records; exclusion: missing data, other eye diseases, or surgeries. The right eye was selected for analysis. Screening myopia was defined as uncorrected distance visual acuity (UDVA) ≥0 logMAR and noncycloplegic spherical equivalent (SE) ≤ -0.5 D; high myopia as SE ≤ -6 D. After quality control, 1,285,609 participants were included (mean age 11.80, SD 3.07; 51.2% male). Prevalence was analyzed by age, sex, school stage, and city GDP level. For prediction, 16,760 participants with 3-year follow-up were selected; 380 developed high myopia, 14,948 did not. After random oversampling (1000 from 380) and undersampling (1000 from 14,948), a balanced dataset of 2000 records was created. Features considered: age, sex, BMI, mean arterial pressure (MAP), sexual maturity, GDP, UDVA, and SE. Permutation importance identified age, UDVA, and SE as key predictors. A random forest algorithm (100 estimators, max depth 19) was trained on 80% of data (5-fold cross-validation) and tested on 20%. External validation used 2547 individuals from Shenzhen Eye Hospital (mean age 8.78, SD 2.39; 47 with high myopia). Performance metrics: accuracy, precision, recall, AUC, and Kolmogorov-Smirnov (K-S) value.
**Key Results:** Overall high myopia prevalence was 4.48% (2019), 4.88% (2020), and 3.17% (2021). Prevalence increased exponentially with age: from 0.43% (≤7 years) to 13.97% (≥18 years) in 2019; similar trends in 2020 and 2021. Annual growth exceeded 1% from age 11 to 17. By school stage, prevalence was 1.25% (primary), 5.60% (middle), and 12.09% (high school) over 3 years (P<.001). Rates declined annually at all stages (e.g., primary: 1.33% to 1.08%; P<.001). Coastal and southern cities had higher prevalence (range 2.60%-5.83%). In 2019, high myopia correlated with GDP (r=0.83, P=.02), but not in 2020 (r=0.05) or 2021 (r=0.41). Age was the most important risk factor (highest permutation importance), followed by MAP and GDP; sexual maturity, BMI, and sex were not significant. The random forest model achieved: accuracy 0.939 (CV), 0.948 (test); recall 0.964 (CV), 0.949 (test); AUC 0.979 (CV), 0.975 (test); K-S 0.885 (CV), 0.895 (test). External validation: accuracy 0.971, AUC 0.957, K-S 0.846. The model outperformed logistic regression (accuracy 0.85), support vector machine (0.88), and k-nearest neighbor (0.88).
**Clinical Implications:** This study demonstrates that real-world vision screening data (age, UDVA, noncycloplegic SE) can accurately predict high myopia onset in children and adolescents using a random forest algorithm. The high prevalence in southern China (up to 13.97% in older teens) underscores the need for early intervention. The model's simplicity (only three variables) makes it practical for large-scale school screening, enabling timely referral for cycloplegic examination and preventive measures. The decline in prevalence from 2019 to 2021 may reflect COVID-19-related changes or improved awareness. The correlation with GDP suggests socioeconomic factors influence myopia risk, possibly through educational pressures. Future development of a smartphone app could facilitate widespread use, contributing to global myopia prevention efforts.