Predictive Performance Of New Artificial Intelligence Intraocular Lens Calculation Formulas
Published 2022
- 40th Congress of the ESCRS
Reference: PP13.01
| Type: ESCRS 2022 - Posters
| DOI:
10.82333/v90m-qt04
Authors:
David Mingo Mingo-Botín* 1
, Francisco Javier Castro-Alonso 2
1Ophthalmology,Hospital Universitario Ramón y Cajal,Madrid,Spain;IMO Madrid - Grupo Miranza,Madrid,Spain, 2Ophthalmology,Hospital de Alcañiz,Alcañiz,Spain;Clínica de Oftalmología Dr FJ Castro,Alcañiz,Spain
Purpose
To evaluate the accuracy of 5 artificial intelligence (AI) based methods for intraocular lens (IOL) power selection (Kane, Hill-RBF 3.0, Hoffer QST, Pearl DGS and Karmona), and compare them to two well established formulas (SRK/T and Barrett Universal II).
Setting
Private practice, Madrid, and Ophthalmology Service, Alcañiz Hospital, Aragón, Spain.
Methods
Retrospective consecutive case series. Patients who had uneventful cataract surgery with implantation of an aspheric IOL (Envista MX60, Bausch&Lomb), and previous Lenstar biometry and Pentacam tomography, were included. Patients with previous corneal refractive surgery, any pathology that could affect the measurements, or poor quality of preoperative examinations were excluded. Prediction error (PE) and its standard deviation (SD), mean (MAE) and median (MedAE) absolute error were calculated for each formula. The percentages of eyes within 0.25 diopters (D), 0.50 D, 0.75 D and 1.00 D of absolute error were also determined.
Results
The study comprised 139 eyes of 124 patients. SD of PE and value of MAE, in order from lowest to highest PE, was Kane (0.318 and 0.243 respectively), Karmona (0.327, 0.262), Pearl-DGS (0.327, 0.247), Barrett (0.333, 0.247), Hill RBF3 (0.337, 0.251), Hoffer-QST (0.347, 0.268) and SRK/T (0.384, 0.299). The lowest MedAE was obtained by Barrett Univesal-II (0.150), followed by Karmona (0.175), Pearl-DGS (0.177), Kane (0.180), Hill-RBF3 (0.183), Hoffer-QST (0.217) and SRK/T (0.250). There were statistically significant differences in MedAE between SRK/T and the other formulas (p<0.03). The highest percentage of eyes with absolute error ≤0.25D was achieved by Pearl-DGS (66.2%), and ≤0.5D and ≤0.75D with Kane (86.3% and 97.8% respectively).
Conclusions
The present study did not find significant differences in the predictive capacity of the current calculation methods using AI. All showed excellent results, exceeding 80% of eyes with absolute error ≤0.5D, although without significantly improving the accuracy of the Barrett Universal-II formula as a reference.