基于随机森林算法的角膜塑形镜近视控制效果预测研究

Prediction of orthokeratology lens in controlling myopia progression by Random Forest algorithm

目的利用随机森林算法建立预测模型预测患者配戴角膜塑形镜的近视控制效果。方法回顾性收集2015年6月至2020年12月在宁波大学附属人民医院眼科中心首次验配角膜塑形镜并随访1年以上的患者,选取验配角膜塑形镜时的年龄、眼轴、角膜平坦曲率、球镜度数以及柱镜度数作为特征,利用随机森林算法对角膜塑形镜控制效果建立预测模型。结果通过变量重要性分数评估发现对于角膜塑形镜控制效果预测准确性贡献度较大的前三位因素为验配角膜塑形镜时的年龄、眼轴及角膜平坦曲率。经多次模型特征调整后,最终使用配戴角膜塑形镜时的年龄、眼轴、角膜平坦曲率、球镜度数及柱镜度数5个特征,利用十折交叉验证的随机森林模型的总体预测准确率为82.61%,特异性为88.9%,敏感性为71.4%,模型值为58.93%,ROC曲线下面积为0.802。结论基于随机森林算法建立的预测模型具有良好的预测效果。

Objective To predict the effect of orthokeratology lens in controlling myopia progression by constructing a model based on Random Forest algorithm.Methods The patients were collected,who carried orthokeratology lens from 2015 to 2020 for the first time in the Affiliated People’s Hospital of Ningbo University.The clinical char acters including age,axial length,corneal flat curvature,sphere power,cy linder power were selected to construct pre diction model by Random Forest algorithm.Results Through the evaluation of variable importance scores,the results showed the top three factors that contributed significantly to the accuracy of predicting the control effect of orthokeratology lenses were age,eye axis,and cormeal flatness curvature when fiting orthokeratology lenses.After multiple model feature adjustments,the overall prediction accuracy of Random Forest algorithm model by ten fold cross validation was 82.61%,with specificity of 88.9%,sensitivity of71.4%,and Kappa value of 58.93%,and the area under the ROC curve was 0.802,based on the age,eye axis,corneal flatness curvature,spherical lens degree,and cylindrical lens degree when wearing orthokeratology lens.Conclusions The model based on Random Forest algorithm can predict the controlling effect of orthokeratology lens well.