Prediction of postoperative vault after implantable collamer lens implantation with deep learning.

Abstract

Aim: To predict the post-operative vault and the suitable size of the implantable collamer lens (ICL) by comparing the performance of multiple artificial intelligence (AI) algorithms.

Methods: A retrospective analysis of 83 patients with 132 eyes was conducted from 2020 to 2023. All patients underwent implantation of EVO-V4C ICLs. ICLs were selected based on STAAR's recommended formula. Postoperative vault values were measured using anterior segment optical coherence tomography (ASOCT). First, feature selection was performed on patients' preoperative examination parameters to identify those most closely related to postoperative vault and incorporate them into the machine learning model. Subsequently, four regression models, namely MLP, XGBoost, RFR, and KNN, were employed to predict the vault, and their predictive performances were compared. The ICL size was set as the prediction target, with the vault and other input features serving as new inputs for predicting the ICL size.

Results: Among all preoperative parameters, 16 parameters were most closely related to postoperative vault and were included in the prediction model. In vault prediction, XGBoost performed the best in the regression model (R²=0.9999), followed by MLP (R²=0.9987) and RFR (R²=0.8982), while the KNN model had the lowest predictive performance (R²=0.3852). XGBoost achieved a prediction accuracy of 99.8%, MLP had a prediction accuracy of 98.9%, while RFR and KNN had accuracies of 87.1% and 57.4%, respectively.

Conclusion: AI effectively predicts postoperative vault and determines ICL size. XGBoost outperforms other machine-learning algorithms tested. Its accurate predictions help ophthalmologists choose the right ICL size, ensuring proper vaulting.