Predictability of the vault after implantable collamer lens implantation using OCT and artificial intelligence in White patient eyes

Abstract

Purpose: To compare the predicted vault using machine learning with the achieved vault using the online manufacturer's nomogram in patients undergoing posterior chamber implantation with an implantable collamer lens (ICL). Setting: Centro Oculistico Bresciano, Brescia, Italy, and I.R.C.C.S.—Bietti Foundation, Rome, Italy. Design: Retrospective multicenter comparison study. Methods: 561 eyes from 300 consecutive patients who underwent ICL placement surgery were included in this study. All preoperative and postoperative measurements were obtained by anterior segment optical coherence tomography (AS-OCT; MS-39). The actual vault was quantitatively measured and compared with the predicted vault using machine learning of AS-OCT metrics. Results: A strong correlation between model predictions and achieved vaulting was detected by random forest regression (RF; R 2 = 0.36), extra tree regression (ET; R 2 = 0.50), and extreme gradient boosting regression (R 2 = 0.39). Conversely, a high residual difference was observed between achieved vaulting values and those predicted by the multilinear regression (R 2 = 0.33) and ridge regression (R 2 = 0.33). ET and RF regressions showed significantly lower mean absolute errors and higher percentages of eyes within ±250 μm of the intended ICL vault compared with the conventional nomogram (94%, 90%, and 72%, respectively; P < .001). ET classifiers achieved an accuracy (percentage of vault in the range of 250 to 750 μm) of up to 98%. Conclusions: Machine learning of preoperative AS-OCT metrics achieved excellent predictability of ICL vault and size, which was significantly higher than the accuracy of the online manufacturer's nomogram, providing the surgeon with a valuable aid for predicting the ICL vault.