目的：探讨机器学习模型预测青少年近视患者配戴角膜塑形镜控制近视效果，并分析影响眼轴增长的关键因素。

方法：采用回顾性临床研究，纳入2021年1月1日至2023年12月1日在本院验配角膜塑形镜并完成6 mo随访的8-17岁青少年近视患者。收集基线一般资料及眼部生物学参数，包括年龄、眼别、睫状肌麻痹验光结果、角膜曲率(K1、K2)及眼轴长度(AL)。以戴镜6 mo后眼轴变化量(ΔAL)为主要结局指标。构建决策树、随机森林、极端梯度提升及神经网络4种机器学习模型，对角膜塑形镜控制效果进行预测，并比较模型性能； 同时采用多元线性回归分析ΔAL的相关影响因素。

结果：本研究共纳入8-17岁青少年近视患者102例159眼，双眼57例，单眼45例，右眼91眼(57.2%)，左眼68眼(42.8%)。其中男56例(54.9%)，女46例(45.1%)； 平均年龄为11.98±1.93 岁，平均球镜度数为-2.28±1.37 D，平均柱镜度数为-0.65±0.36 D，AL为24.60±0.82 mm，K1值为42.55±1.39 D，K2值为43.58±1.44 D。戴镜6 mo后AL为24.68±0.59 mm，平均球镜度数为-2.45±1.46 D，柱镜度数为-0.68±0.38 D，K1值为42.52±1.33 D，K2值为43.61±1.51 D，ΔAL为0.08±0.17 mm。4 种机器学习模型预测角膜塑形镜控制 AL 效果的性能对比显示，决策树模型预测效果最优，其 MAE=0.11、MSE=0.02、R²=0.44。决策树模型输出的特征重要性排序前三名分别为年龄、戴镜前AL、戴镜前角膜水平曲率(K1)。多元线性回归分析显示，年龄与ΔAL呈负相关(β=-0.03，P<0.001)，而其余变量与ΔAL无显著相关性(均P>0.05)。

结论：机器学习模型可有效预测角膜塑形镜对青少年近视患者眼轴增长的控制效果，其中决策树模型表现最佳。年龄是影响角膜塑形镜疗效的关键因素，戴镜前AL及角膜曲率亦具有一定预测价值。本研究为青少年近视个体化防控提供了新的数据支持。

METHODS: This retrospective study included myopic subjects aged 8-17 y who were fitted with OK lenses and completed a 6 mo follow-up between January 1, 2021 and December 1, 2023. Baseline demographic and ocular biometric parameters were collected, including age, eye laterality, cycloplegic refraction, corneal curvature(K1 and K2)and AL. AL change(ΔAL)at 6 mo after lens wearing was defined as the primary outcome. Four machine learning models—decision tree, random forest, extreme gradient boosting(XGBoost), and neural network—were developed to predict treatment efficacy, and their predictive performance was compared. Multiple linear regression analysis was performed to identify relevant factors associated with ΔAL.

RESULTS: A total of 102 myopic patients(159 eyes)aged 8-17 y were included in this study, including 57 bilateral eyes and 45 unilateral cases. Among them, 56 were boys(54.9%)and 46 were girls(45.1%). The mean age was 11.98±1.93 y. At baseline, the mean spherical error was -2.28±1.37 D, the mean cylindrical power was -0.65±0.36 D, and AL was 24.60±0.82 mm, the mean flat keratometry(K1)was 42.55±1.39 D, and the mean steep keratometry(K2)was 43.58±1.44 D.After 6 mo of lens wear, the AL increased to 24.68±0.59 mm, The mean spherical power was -2.45±1.46 D, the cylindrical power was -0.68±0.38 D, the K1 value was 42.52±1.33 D, and the K2 value was 43.61±1.51 D. The mean ΔAL was 0.08±0.17 mm.The performance comparison of four machine learning models in predicting the efficacy of OK on AL control showed that the decision tree model achieved the best predictive performance(MAE=0.11, MSE=0.02, R²=0.44). Feature importance analysis indicated that age, baseline AL and K1 were the top three predictors. Multiple linear regression showed that age was negatively associated with ΔAL(β=-0.03, P<0.001), while other variables were not significantly associated(all P>0.05).

CONCLUSION: Machine learning models can effectively predict the efficacy of OK in controlling axial elongation in children. The decision tree model shows the best performance. Age is a key factor affecting the therapeutic effect of OK, while baseline AL and corneal curvature also contribute to prediction.This study provides new data support for individualized myopia prevention and control strategies in the pediatric population.