Enhancing Intraocular Lens Power Calculation With Bayesian Statistics Compared To Different Artificial Intelligent Approaches
Published 2024
- 42nd Congress of the ESCRS
Reference: FP06.12
| Type: Free paper
| DOI:
10.82333/4d8j-5c89
Authors:
Leon Pomberger* 1
, Lisa Tasch 1
, Cristoph Mayer-Xanthaki 2
, Haidar Khalil 1
, Klemens Waser 1
, Matthias Bolz 1
, Nino Hirnschall 1
1Ophthalmology and Optometry,Johannes Kepler University,Linz,Austria;Ophthalmology and Optometry,Kepler University Clinic,Linz,Austria, 2Ophthalmology and Optometry, Medical University of Graz,Graz,Austria
Purpose
To develop a Bayesian prediction model for intraocular lens (IOL) power calculation and to compare this model with different machine learning approaches.
Setting
Department of Ophthalmology and Optometry, Kepler University Clinic and Johannes Kepler University, Linz, Austria; Department of Ophthalmology, Medical University Graz, Austria
Methods
This multi-center study included 400 patients that underwent cataract surgery. All eyes received a swept-source OCT biometry (IOL Master 700, Carl Zeiss Meditec AG, Germany) followed by implantation of a monofocal hydrophobic open loop haptic IOL (ZCB00, or ICB00, DIU, Johnson&Johnson, USA). Study examination was performed minimum 4 weeks and maximum 24 months after surgery and included optical biometry, autorefraction as well as subjective refraction. The primary focus of the study was on the spherical equivalent obtained from subjective refraction. A Bayesian prediction model for IOL power calculation was introduced. Furthermore, 4 different machine learning models were developed and all different prediction models were compared.
Results
In total, 400 eyes of 400 patients were included at 2 recruitment centers. In a preliminary data set the average axial eye length was 22.66 mm (SD: 2,2), the average anterior chamber depth was 3.07 mm (SD: 0.44) and the average lens thickness was 4.58mm (SD: 0.51). After constant optimization for different IOL power calculation formulae the mean absolute refractive error (MAE) was found to be between 0.28 D and 0.48 D. A significant improvement was found when comparing the newly developed Bayesian statistics approach, although the number of outlayers could not be reduced significantly. The comparison between the Bayesian prediction model and the machine learning approaches will be presented at the meeting.
Conclusions
Bayesian statistics can improve the post-operative refractive outcome and reduce, but not eliminate the number of refractive surprises.