Exploring A Machine Learning Model Predicting Myopic Regression Following Small Incision Lenticule Extraction And Femtosecond Laser-Assisted In Situ Keratomileusis
Published 2025 - 43rd Congress of the ESCRS
Reference: FP02.13 | Type: Free paper | DOI: 10.82333/h7cn-x841
Authors: Gabriele Gallo Afflitto* 1 , Francesco Aiello 2 , Pier Luigi Surico 3 , Davina A. Malek 4 , Tommaso Mori 5 , Swarup S. Swaminathan 4 , Vincenzo Maurino 6 , Carlo Nucci 2
1Moorfields Eye Hospital NHS Foundation Trust,London,United Kingdom;Università di Roma "Tor Vergata",Roma,Italy, 2Università di Roma "Tor Vergata",Roma,Italy, 3Department of Organs of Sense,Università di Roma "La Sapienza",Roma,Italy, 4Bascom Palmer Eye Institute, University of Miami Miller School of Medicine,Miami,United States, 5Moores Cancer Center,University of California San Diego,La Jolla,United States, 6Moorfields Eye Hospital NHS Foundation Trust,London,United Kingdom
Purpose
To assess the effectiveness of the machine learning model in predicting myopic regression in patients undergoing Small Incision Lenticule Extraction (SMILE) and Femtosecond Laser-Assisted in Situ Keratomileusis (FS-LASIK)
Setting
The retrospective cohort study was conducted at Beijing AierIntech Eye Hospital.The study involved patients with myopia and astigmatism who underwent SMILE or FS-LASIK between January 2010 and March 2021.
Methods
The participants were divided into training and validation sets employing the chronological data-splitting method: follow-up conducted over 3 months to 2 years. A myopic regression model was developed using Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Gradient Boosting Machine (GBM) after the predictors were evaluated with Univariate Cox proportional hazards (Cox PH) analysis. Models are verified on the training set and validation set. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was evaluated for the predictive models' performance. Decision Curve Analysis (DCA) is utilised to assess clinical utility.
Results
12182 eyes were included, with 8,513 selected for the training set and 3,669 for validation. After a Cox PH analysis and considering clinical significance to identify predictors, the training and validation sets were employed to assess the ROC and AUC for the RF, Extreme XGBoost, and GBM models. GBM exhibited superior results with a ROC-AUC of 0.734 (95% CI, 0.715-0.754) in the training set and 0.655 (95% CI, 0.623- 0.687) in the validation set, respectively. The eight key predictors include preoperative spherical diopter, anterior chamber depth, central corneal thickness, age, higher-order aberrations within a 5mm pupil, Kmax/Kmin (the corneal curvature in the steep/flat meridians), and ablation depth.
Conclusions
When comparing machine learning algorithms, GBM emerged as the most efficient model for estimating high-risk myopic regression patients following SMILE and FS-LASIK based on preoperative and operative data. During the preoperative consultation, the surgeon could evaluate the patient's risk of myopic regression alongside the available surgical options. The patient would benefit from the surgeon's assessment of the risk of myopic regression and the available surgical options.