Predicting Phakic Intraocular Lens Misalignment For Myopia By Machine Learning Models

 To predict postoperative phakic intraocular lens rotation by machine learning models and big data analytics.

This study was conducted at Eye, Ear, Nose and Throat Hospital, Fudan University,Shanghai, China

 This study included 642 eyes of 371 patients (mean age 26.56±5.42 years) who underwent Toric implantable collamer lens(TICL) surgery. We trained four regression models to quantitatively predict rotation degree, and the Random Forest to qualitatively predict whether the TICL rotation degree ＞10°, whether misalignment cause residual astigmatism error or visual acuity loss, and the need for realignment surgery.  Permutation importance and impurity-based feature importance were used to explore factors for predicting misaligment degree. Regression models were evaluated using RMSE and R², and classification models were evaluated using accuracy and area under the curve(AUC). 

The mean absolute postoperative rotation degree was 6.14 ± 9.63°. Among all 642 eyes,181 eyes (28.15%) had residual astigmatism error over 0.50D and 62 eyes (9.64%) loss vision due to TICL misalignment, and 20 eyes (3.11%) underwent realignment surgery. Preoperative manifest sphere, WTW, ACV, horizontal STS, vertical STS, and STS-WTW were the most important factors. In test set, XGBoost performed best for predicting postoperative rotation degree (RMSE=5.10,R²=0.39). For predicting whether postoperative rotation over 10°, misalignment-related residual astigmatism error, vision loss and the need for realignment surgery, Random Forest had an accuracy of 0.9020,0.7761,0.9030, and 0.9776, with an AUC of 0.7181,0.5976,0.6619, and 0.8532.

By inputting preoperative sphere, WTW, ACV, horizontal STS, vertical STS, STS-WTW, and other ocular biological parameters, the machine learning model can predict postoperative rotation degree, which can improve the refractive and visual predictability after TICL implantation.