Development And Validation Of The Keratoconus Artificial Intelligence Index (Kai-Index).
Published 2026 - 44th Congress of the ESCRS
Reference: FP08.03 | Type: Free paper | DOI: 10.82333/b5hr-fq57
Authors: Dr Riccardo Vinciguerra* 1 , Mr Francesco Versaci 2 , A/Prof. Jorge Alio Del Barrio 3 , Mr Robert Herber 4 , Dr Emilio Torres 5 , A/Prof. Paolo Vinciguerra 6 , A/Prof. Farhad Hafezi 7 , Dr Juan Arbelaez 8 , Dr Frederick Raiskup 4 , A/Prof. Jorge Luis Alio 3 , A/Prof. Mario Romano 6 , Dr Renato Ambrosio 9
1Humanitas University,Rozzano,Italy;Humanitas University,Rozzano,Italy, 2CSO Italia,Firenze,Italy, 3Universidad Miguel Hernández,Alicante,Spain, 4University Hospital Carl Gustav Carus,Dresden,Germany, 5Elza Institute,Zurich,Switzerland, 6Humanitas University,Rozzano,Italy, 7ELZA Institute,Zurich,Switzerland, 8Muscat Eye Laser Center,Muscat,Oman, 9Department of Ophthalmology, Federal University the State of Rio de Janeiro,Rio de Janeiro,Brazil
Purpose
To develop and validate the Keratoconus Artificial Intelligence Index (KAI Index), an artificial intelligence–based metric integrating corneal tomography and epithelial thickness mapping for the early detection of keratoconus.
Setting
A total of 4,200 eyes (1,992 normal, 1,793 keratoconus, and 415 very asymmetric ectasia with normal tomography [VAE-NT]) recruited from five international centers
Methods
All eyes underwent imaging with the MS-39 tomography system. Eyes were classified as normal, bilateral keratoconus, or VAE-NT using clinical, topographic, and objective criteria. One eye per subject was included, except for VAE-NT cases. Quantitative tomographic and epithelial thickness parameters were analyzed. Feature selection was performed using receiver operating characteristic (ROC) analysis and LASSO regression. Logistic regression, support vector machine, and random forest classifiers were trained using subject-level stratified splits. An external validation cohort was used. The resulting KAI Index was compared with established indices (BAD-D, CBI, TBIv2, and PRFI) using ROC analysis and DeLong testing.
Results
Random forest demonstrated the best performance and was selected as the final model. In the internal test set, the KAI Index achieved an AUC of 1.000 for normal versus keratoconus and 0.971 for normal versus VAE-NT.
In the external validation cohort, the AUC was 1.000 for normal versus keratoconus, comparable to BAD-D (0.995), PRFI (0.987), and TBIv2 (0.994), and higher than CBI (0.971, p=0.0341). For discrimination between normal and VAE-NT eyes, the KAI Index achieved an AUC of 0.878, exceeding TBIv2 (0.844 while not significantly, p=0.25), CBI (0.771), PRFI (0.752), and BAD-D (0.708). Differences were statistically significant for BAD-D (p<0.0001), PRFI (p = 0.0002) and CBI (p=0.02).
Conclusions
The KAI Index is a novel, artificial intelligence–driven, epithelial-integrated tomographic metric that accurately detects keratoconus and very early ectatic disease. Its robust performance in internal and external validation supports its potential clinical utility for refractive surgery screening and early risk stratification for ectasia.