The 1000 Eyes Study: Introducing Bayes Statistics And Artificial Intelligence To Iol Power Calculation

Published 2023 - 41st Congress of the ESCRS

Reference: FP05.08 | Type: Free paper | DOI: 10.82333/hq8c-5s47

Authors: Lisa Tasch* ¹ , Christoph Mayer-Xanthaki ² , Leon Pomberger ¹ , Haidar Khalil ¹ , Matthias Bolz ¹ , Nino Hirnschall ¹

¹Ophthalmology and Optometry,Johannes Kepler University,Linz,Austria;Ophthalmology and Optometry,Kepler University Clinic,Linz,Austria, ²Ophthalmology,Medical University Graz,Graz,Austria

Purpose

To train regression models for intraocular lens (IOL) power calculations using several machine learning approaches in order to minimize the refractive error after cataract surgery.

Setting

Department of Ophthalmology and Optometry, Johannes Kepler University, Linz, Austria; Department of Ophthalmology and Optometry, Kepler University Clinic, Linz, Austria; Department of Ophthalmology, Medical University Graz, Austria

Methods

This two-center study included patients that underwent cataract surgery between July 2021 and January 2022. All eyes received a swept-source OCT biometry (IOL Master 700) followed by implantation of a monofocal hydrophobic open loop haptic IOL. Study examination was performed minimum 4 weeks and maximum 24 months after surgery and included optical biometry, autorefraction as well as subjective refraction. Main parameter for the machine learning approached was the spherical equivalent deriving from subjective refraction. Next to classical random forest plots, partial least squares regression (PLSR) and Bayes statistics were used to predict the mentioned primary endpoint parameter.

Results

In total, 1000 eyes of 1000 patients were included in this trial. After constant optimisation for different traditional IOL power calculation formulae the mean absolute refractive error (MAE) was found to be between 0.33 D and 0.48 D. A significant improvement was found when comparing PLSR to traditional formulae (p<0.01 for all), however no significant difference in MAE was found when comparing PLSR to classical random forest plot (0.24±0.24 vs. 0.28±0.27 respectively, p=0.684). Furthermore, a concept based on Bayes statistics will be presented at the ESCRS meeting.

Conclusions

PLSR can improve the post-operative refractive outcome and decrease the number of refractive surprises. However, not all cases of refractive surprises can be avoided by this method.