ESCRS - PO550 - Keratoconus Progression Prediction Using Different Ai Models And A Novel Approach For Segmentation Of Epithelial Mapping

Keratoconus Progression Prediction Using Different Ai Models And A Novel Approach For Segmentation Of Epithelial Mapping

Published 2025 - 43rd Congress of the ESCRS

Reference: PO550 | Type: Free paper | DOI: 10.82333/fg7b-wk19

Authors: Magdalena Niestrata* 1 , James Jackson 2 , Shehnaz Bazeer 3 , Jose Galvez-Olortegui 4 , Bruce Allan 3

1Ophthalmology,East Sussex and North Essex Foundation Trust,London,United Kingdom, 2Governance and Data,University of East London,London,United Kingdom, 3Ophthalmology,Moorfields Eye Hospital,London,United Kingdom, 4Ophthalmology,Scientia Clinical and Epidemiological Research Institute,Trujillo,Peru

Purpose

Aim of this study was to predict the progression of keratoconus using multiple artificial intelligence prediction models trained with data of 12 month follow ups and a segmentation approach for epithelial mapping.

Setting

Kepler University Clinic Linz, MCIII, Department of Ophthalmology and Optometry and Johannes Kepler University.

Methods

Patient data was collected from our Keratoconus outpatient clinic in a prospective fashion. In all cases a spectral domain OCT (MS-39, CSO, Italy) was available at least over a time period of 12 months. Progression was defined as an increase of Kmax of ≥1 Diopter. Multiple binomial and regressive AI Models were trained with 16 most predictive Keratoconus parameters. A novel  approach for segmentation of epithelial mapping was applied using MedSAM.

Results

In total 200 eyes of 100 patients were included in this study. Mean age was 33.13 years (SD 13.56; median: 30; range 8 to 74). Preliminary results showed a Kmax of 52.47 D (SD: 7.85) and a mean tomographic astigmatism of 3.08 D (SD: 2.12). Thinnest pachymetry was on average 470.64μm (SD: 65.10).  Of all the different models the binomial random forest machine learning model was found to be promising as well as a bayesian model especially in combination with segmentation of epithelial mapping.

Conclusions

Prediction of keratoconus progression using random forest plots and Bayesian statistics in combination with the segmentation of epithelial mapping is a novel powerful tool that could be used to detect subclinical keratoconus for early intervention.