Posters

A new model for IOL calculation in eyes with prior corneal refractive surgery developed using machine learning

Poster Details

First Author: M.Dubois FRANCE

Co Author(s):    G. Debellemanière   S. Moran   L. Arnould   A. Saad   D. Gatinel        

Abstract Details

Purpose:

To evaluate the accuracy of  a new IOL power calculation method developed using machine learning (XGB), by comparing results with no history formulas from the ASCRS calculator (Shammas, Haigis L, Barrett True K, Double-K Holladay ) in the prediction of postoperative refraction for eyes with previous corneal refractive surgery

Setting:

Retrospective single-center consecutive case series carried out at the Rothschild Foundation, Paris

Methods:

Patients undergoing cataract surgery between january 2017 and december 2018  were analyzed.  Preoperative biometry was obtained using the IOLMaster 700. Eyes that have undergone refractive surgery were included. The criteria for exclusion were eyes with corneal pathology, incomplete data,  or other issues with potential to influence postoperative refraction. 80% of data set was used to train a machine learning algorithm and the remaining 20% was left untouched to compare the performance of each formula with respect to the real obtained spherical equivalent. The analysis was focused on eyes with previous corneal refractive surgery  among the validation set in this study.

Results:

2560 eyes were part of the training set. We had 14 radial keratotomy (RK), 13 photorefractive keratectomy (PKR) and 21 Lasik. The Median Absolute Error (MAE) was 0,49; 0,78; 0,84 and 1,12 after PKR respectively for XGB, Haigis L, Barrett True K (BTK) and Shammas formulas. The MAE was 0,55 ; 0,83; 0,88 and 1,16 after Lasik respectively for XGB, Haigis L, BTK and Shammas formulas with a statistically significant superiority for the XGB model (p<0,05).  No significant difference was found in post  RK group

Conclusions:

The XGB model was superior to Shammas formula in case of PKR, and to all formulas except Haigis L in case of Lasik. Data-based formulas provide a  powerful new method of IOL power calculation, and are of particular interest in "atypical" eyes. Application of this formula to larger datasets in the wider population, will in turn continue to improve the performance of such formulas

Financial Disclosure:

None