Accuracy Of A New Intraocular Lens Power Calculation Method Based On Artificial Intelligence
Published 2022 - 40th Congress of the ESCRS
Reference: FPT04.03 | Type: Free paper | DOI: 10.82333/a49n-8h37
Authors: Carlos Palomino Bautista* 1 , David Carmona 1 , Alfredo Castillo 1 , Marta Romero 1 , Ruben Sanchez Jean 2
1Hospital Universitario Quironsalud Madrid,Pozuelo de Alarcón,Spain, 2Hospital Clínico San Carlos,Madrid,Spain
Purpose
Setting
Methods
A sample of 481 eyes undergoing uneventful cataract surgery by 4 surgeons with different types of intraocular implants was collected. All eyes were measured preoperatively with IOL Master® 700 and Pentacam® HR for corneal posterior curvature.
The dataset was prepared and analyzed for variable selection by eliminating correlations that produced collinearity. The sample was randomized and subsequently split into two parts with a ratio 80 - 20: training and test.
The regression models were implemented in raw form using Machine Learning techniques. Subsequently, they were optimized and hyperparameterized to improve and enhance predictability. Stacking techniques were used to assemble the best final models, obtaining a final model called Karmona.
Results
Karmona showed the best performance on the validation sample corresponding to 20% of the original dataset (n = 95), obtaining a MAE ± SD (MedAE) of 0.23 ± 0.20 D (0.19 D) with 92% of eyes included in the postoperative refractive interval of ±0.50 D, followed by PEARL-DGS with 0.27 ± 0.25 D (0.22 D) with 88% of eyes in ±0.50 D, with no statistically significant differences.
Conclusions
Karmona has shown better results than the rest of the formulas with which it has been compared, for the population studied in this work.