Predicting Vault and Size of Posterior Chamber Phakic Intraocular Lens Using Sulcus to Sulcus-Optimized Artificial Intelligence Technology.

Purpose: To investigate the accuracy of posterior chamber phakic intraocular lens (PIOL) vault and size prediction models based on sulcus to sulcus (STS) optimized artificial intelligence and big data analysis technology.

Design: Big data and artificial intelligence prediction model.

Methods: We included 5873 eyes with posterior chamber PIOL implantation, and the postoperative vault was measured using an anterior segment analyzer (Pentacam AXL) 1 month postoperatively. A random forest regression model and classification model were used to predict the postoperative vault and PIOL size. The postoperative vault and PIOL size were set as output features; other vault-related eye parameters were set as input features. The influence of white to white (WTW), horizontal sulcus to sulcus (STS), and vertical STS on predicting postoperative vault and PIOL size was analyzed and compared.

Results: The mean preoperative WTW diameter was 11.64 ± 0.37 mm, the mean horizontal STS diameter was 11.85 ± 0.47 mm, and the mean vertical STS diameter was 12.39 ± 0.52 mm. In the regression model for numerical prediction of the vault, the combination of WTW, horizontal STS, and vertical STS was the most optimal for vault prediction (R = 0.3091, root mean square error [RMSE] = 0.1705); solely relying on WTW was the least optimal (R = 0.2849, RMSE = 0.1735). Among the models for classification prediction of the vault, the combination of WTW, horizontal STS, and vertical STS was the most accurate (accuracy, 0.6302; mean area under the curve, 0.8008; and mean precision recall rate, 0.6940). Moreover, the combination of WTW, horizontal STS, and vertical STS exhibited the highest accuracy for classification prediction of PIOL size (accuracy, 0.8170; mean area under the curve, 0.9540; and mean precision recall rate, 0.8864). Whether in the regression prediction models of vault values or in the classification prediction models of vault and PIOL size, the accuracy of STS optimized model was significantly improved compared with the traditional WTW model (P < .001).

Conclusion: Artificial intelligence combined with STS optimization contributes to the accuracy of PIOL size and vault prediction models. The random forest machine-learning model optimized by STS is superior to the traditional WTW model.