The Accuracy Of Intraocular Lens Power Calculation Formulas Based On Artificial Intelligence In Eyes With Myopic Axial Length: A Systematic Review And Meta-Analysis
Published 2025 - 43rd Congress of the ESCRS
Reference: PO892 | Type: Poster | DOI: 10.82333/r3p1-5e13
Authors: Maria Agapi Keventzidou* 1 , Asimina Mataftsi 2 , Georgios Katsaras 3 , Georgios Lavasidis 4 , Ioannis Tsinopoulos 2
1Ophthalmology,Master of Science (MSc) “Ocular Surgery”, School of Medicine, Aristotle University of Thessaloniki,Thessaloniki,Greece;Department of Ophthalmology, Children’s Hospital “Agia Sofia”,Athens,Greece;2nd Department of Ophthalmology, School of Medicine, Aristotle University of Thessaloniki, Papageorgiou General Hospital,Thessaloniki,Greece, 2Ophthalmology,Master of Science (MSc) “Ocular Surgery”, School of Medicine, Aristotle University of Thessaloniki,Thessaloniki,Greece;2nd Department of Ophthalmology, School of Medicine, Aristotle University of Thessaloniki, Papageorgiou General Hospital,Thessaloniki,Greece, 3Ophthalmology,Master of Science (MSc) “Ocular Surgery”, School of Medicine, Aristotle University of Thessaloniki,Thessaloniki,Greece, 4Ophthalmology,Master of Science (MSc) “Ocular Surgery”, School of Medicine, Aristotle University of Thessaloniki,Thessaloniki,Greece; Department of Ophthalmology, “Elpis” General Hospital,Athens,Greece
Purpose
To investigate the accuracy of formulas based on artificial intelligence (AI) compared with 3rd, 4th and newer generation formulas in eyes with medium-long (24.5-26.0mm) and long axial length (AL) (>26.0mm).
Setting
This systematic review and meta-analysis of observational studies was conducted at the 2nd Department of Ophthalmology, Aristotle University of Thessaloniki School of Medicine.
Methods
The literature was searched until 24/1/24 for studies comparing 8 formulas AI-based (Kane, Ladas Super Formula-LSF, Hill-RBF 2.0, Hill-RBF 3.0, XGBoost, K6, FullMonte, Pearl-DGS) with 8 3rd, 4th and newer generation formulas with Wang-Koch adjustment if available (Barrett Universal II-BUII, SRK/T, Holladay 1, Holladay 2, Hoffer Q, Haigis, Olsen, Emmetropia Verifying Optical-EVO) in eyes with AL>24.5mm. Primary outcomes were the percentage of eyes with refractive prediction error (PE) within ±0.25D, ±0.5D and ±1.0D. Secondary outcomes included Mean Absolute Error (MAE) and Median Absolute Error (MedAE). Subgroup analyses were based on optical biometry and axial length. A sensitivity analysis excluded studies with high risk of bias.
Results
A total of 12 studies involving 1959 eyes with myopic AL>24.5mm met the inclusion criteria. Kane and Hill-RBF exhibited higher or equivalent prediction accuracy than conventional formulas regarding percentage of eyes with PE within ±0.25D, ±0.5D, ±1.0D and had the lowest MAE and MedAE values. LSF was comparable in accuracy to traditional formulas. However, LSF was outperformed by BUII in PE ±0.25D (p=0.01), ±0.5D (p=0.0008) and by EVO in PE within ±0.5D (p=0.02). XGBoost formula was superior to BUII for PE within ±0.25D, ±0.5D, ±1.0D and exhibited significantly more accurate outcomes for eyes with AL>30mm (p=0.0001). FullMonte was found to be inferior to BUII in terms of PE ±0.25D (p=0.009) and ±0.5D (p=0.0008).
Conclusions
AI-driven intraocular lens power calculation formulas present enhanced predictive accuracy over more traditional methods.The foremost AI-based formulas for predicting IOL power for AL>24.5mm are Kane, Hill-RBF and XGBoost. They were equivalent or superior to 3rd, 4th, newer generations and even to formulas with Wang-Koch adjustments. Nevertheless, larger-scale studies with broader geographic representation are essential to validate the positive outcomes of AI-based models in cataract surgery.