ESCRS - POREF214 - Artificial Intelligence Algorithms For Enhancing Photorefractive Keratectomy Safety And Efficiency

Artificial Intelligence Algorithms For Enhancing Photorefractive Keratectomy Safety And Efficiency

Published 2026 - 44th Congress of the ESCRS

Reference: POREF214 | Type: Free paper | DOI: 10.82333/whh7-et68

Authors: A/Prof. Vincent Borderie* 1 , Mr Nassim Louissi 1 , Dr Benjamin Memmi 1 , Dr Juliette Knoeri 1 , Mrs Cristina Georgeon 1 , A/Prof. Nacim Bouheraoua 1 , Dr Anatole Chessel 2

1Ophthalmology,Hôpital national d'ophtalmogie des 15-20,Paris,France, 2Laboratory for Optics and Biosciences ,École Polytechnique,Palaiseau,France

Purpose

To develop artificial intelligence-based models capable of predicting the achieved stromal ablation depth (AD), postoperative spherical and cylindrical corrections, and the laser-induced change in corneal curvature to enhance the safety and efficiency of photorefractive keratectomy (PRK).

Setting

French national eye hospital (Hôpital national des 15-20), Paris, France.

Methods

Design: retrospective cohort study including 311 eyes with PRK with at least three months of follow-up, pre- and postoperative optical coherence tomography (OCT), and corneal topography. Preoperative subjective spherical equivalent (SE) and astigmatism ranged, respectively,  from -8.00D to +4.00D and from 0.00D to 5.00D. Main outcome measures: 3-month measured AD, change in subjective spherical equivalent (SE), surgically induced astigmatism cylinder (SIA), and change in mean simulated keratometry (mean SimK). The collected data were divided into a training set (70%) and a test set (30%). Artificial intelligence frameworks were used to predict the outcome measures.

Results

Compared with the laser nomogram, machine learning and deep learning models provided better prediction of AD (r2: model, 0.91 versus laser, 0.84; Mean Absolute Error: 6.79D versus 9.51D), change in SE (r2: model, 0.97 versus  0.96; MAE: 0.36D versus 0.42D), SIA (r2: model, 0.84 versus 0.69; MAE: 0.29D versus 0.37D), and change in mean SimK (r2: model, 0.96 versus 0.95; MAE: 0.34D versus 0.40D) in the test set. An r2 of 0.99 was achieved when full preoperative OCT and corneal topography data were used as inputs to the models.

Conclusions

Artificial intelligence models show promising potential to improve PRK safety and efficiency by refining the laser nomogram.