Predicting intraocular lens tilt using a machine learning concept.

Abstract

Purpose: To use a combination of partial least squares regression and a machine learning approach to predict intraocular lens (IOL) tilt using preoperative biometry data.

Setting: Kepler University Clinic Linz, Linz, Austria.

Design: Prospective single-center study.

Methods: Optical coherence tomography, autorefraction, and subjective refraction were performed at baseline and 8 weeks after cataract surgery. In analysis I, only 1 eye per patient was included and a tilt prediction model was generated. In analysis II, a pairwise comparison between right and left eyes was performed.

Results: In analysis I, 50 eyes of 50 patients were analyzed. Difference in amount, orientation, and vector from preoperative to postoperative lens tilt was -0.13 degrees, 2.14 degrees, and 1.20 degrees, respectively. A high predictive power (variable importance for projection [VIP]) for postoperative tilt prediction was found for preoperative tilt (VIP = 2.2), pupil decentration (VIP = 1.5), lens thickness (VIP = 1.1), axial eye length (VIP = 0.9), and preoperative lens decentration (VIP = 0.8). These variables were applied to a machine learning algorithm resulting in an out of bag score of 0.92 degrees. In analysis II, 76 eyes of 38 patients were included. The difference of preoperative to postoperative IOL tilt of right and left eyes of the same individual was statistically relevant.

Conclusions: Postoperative IOL tilt showed excellent predictability using preoperative biometry data and a combination of partial least squares regression and a machine learning algorithm. Preoperative lens tilt, pupil decentration, lens thickness, axial eye length, and preoperative lens decentration were found to be the most relevant parameters for this prediction model.