Liu Pan, Gu Xiaochen, Jiao Yexuan, Ye Xinqi, Zhou Yu-Hang, Wang Xinlin, Zhou Yongjin, Shao Zhengbo
Department of Ophthalmology, The Second Affiliated Hospital of Harbin Medical University, Harbin, 150086, China.
Future Medical Laboratory, The Second Affiliated Hospital of Harbin Medical University, Harbin, 150086, China.
Int Ophthalmol. 2025 May 5;45(1):175. doi: 10.1007/s10792-025-03520-7.
To predict nomogram for small incision lenticule extraction (SMILE) using machine learning technology and preoperative clinical data.
A total of 1025 eyes with postoperative spherical equivalent within ± 0.50D after SMILE were included in this study. The XGBoost, gradient boosting regression (GBR), random forest (RF), LightGBM, linear regression (LR) and support vector regression (SVR) were applied to predict the nomogram. The performance of six machine learning methods was assessed by calculating the root mean absolute error (RMSE) and the mean absolute error (MAE). Four junior residents were selected to design the nomogram based on preoperative clinical data in testing set, and were compared with the machine learning models by calculating the accuracy of eyes within three specific thresholds (± 0.05D, ± 0.15D, ± 0.25D).
The actual nomogram was not significantly different from the nomogram predicted machine learning methods (P > 0.05). The RMSE of six models ranged from 0.075 to 0.110, and MAE were 0.055 to 0.085 on nomogram prediction. The XGBoost provided significantly higher accuracy within 0.05 to 0.25 D than the SVR and junior residents (McNemar test, P < 0.001). However, there were no statistically significant differences in accuracy within 0.05 to 0.25 D that the XGBoost, GBR, RF, LightGBM, and LR achieved (P > 0.05).
Machine learning of the preoperative clinical data could accurately predict nomogram for SMILE. The machine learning methods may assist the refractive surgeons and shorten the learning curve of junior residents while making the nomogram adjustment.
利用机器学习技术和术前临床数据预测小切口透镜切除术（SMILE）的列线图。
本研究纳入了1025只在SMILE术后等效球镜度在±0.50D以内的眼睛。应用XGBoost、梯度提升回归（GBR）、随机森林（RF）、LightGBM、线性回归（LR）和支持向量回归（SVR）来预测列线图。通过计算均方根误差（RMSE）和平均绝对误差（MAE）来评估六种机器学习方法的性能。选择四名初级住院医师根据测试集中的术前临床数据设计列线图，并通过计算在三个特定阈值（±0.05D、±0.15D、±0.25D）内眼睛的准确性与机器学习模型进行比较。
实际列线图与机器学习方法预测的列线图无显著差异（P>0.05）。六种模型在列线图预测中的RMSE范围为0.075至0.110，MAE为0.055至0.085。在0.05至0.25D范围内，XGBoost的准确性显著高于SVR和初级住院医师（McNemar检验，P<0.001）。然而，XGBoost、GBR、RF、LightGBM和LR在0.05至0.25D范围内的准确性无统计学显著差异（P>0.05）。
术前临床数据的机器学习可以准确预测SMILE的列线图。机器学习方法可以帮助屈光手术医生，缩短初级住院医师的学习曲线，同时进行列线图调整。
