Artificial Intelligence-Assisted And Ciliary Sulcus Diameter-Optimized Evo-Icl Size Selection And Vault Prediction For Myopia Correction

 The purpose of the study is to optimize the ICL size selection and vault prediction by artificial intelligence (AI) and big data analytics.

The medical records of patients who underwent EVO-ICL implantation at the Eye and ENT Hospital, Fudan University (From Jan. 1st, 2015 to Dec. 31st 2021) were included in this retrospective study. This study was completed in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Eye and ENT Hospital, Fudan University (approval number 2021018).

In total, 5873 eyes of 5873 patients who underwent ICL implantation were included. The vault was measured using an anterior segment analyzer (Pentacam AXL) at postoperative 1 month. Random forest regression and classification models were used for vault prediction and ICL sizing. The vault and the selected ICL size were set as output features. The values of flat keratometry, steep keratometry, anterior chamber depth, anterior chamber angle, pupil size, central corneal thickness, horizontal corneal diameter (WTW), axial length, horizontal ciliary sulcus diameter (hSTS) and vertical ciliary sulcus diameter (vSTS) were set as input features. The effects of WTW, hSTS, and vSTS on vault predicting and ICL sizing were evaluated.

In the regression model for numerical prediction of the vault, the combination of WTW, hSTS, and vSTS was the most optimal for vault prediction (R²=0.3091, root mean square error [RMSE]= 0.1705) ; solely relying on WTW was the least optimal (R²=0.2849, RMSE=0.1735). Among the models for classifying the vault, the combination of WTW, hSTS, and vSTS was the most accurate, with accuracy (ACC) of 0.6302, area under the curve (AUC) of 0.8008. Moreover, the combination of WTW, hSTS, and vSTS yielded the highest accuracy for ICL sizing (ACC=0.8170; AUC=0.9540). The ACC of STS-optimized model for ICL sizing was significantly improved when compared with the traditional WTW model (P<0.001).

AI is applicable for predicting vault and ICL size. Optimal prediction accuracy can be obtained through a combination of WTW, hSTS, and vSTS. The random forest machine-learning model optimized by STS for ICL sizing is more accurate than the traditional WTW model.