Development And Validation Of The Keratoconus Artificial Intelligence Index (Kai-Index).

Optical Coherence Tomography (OCT) is widely used for anterior segment imaging due to its high resolution and ability to measure epithelial thickness and track the posterior corneal surface. However, its role in detecting corneal ectasia, particularly keratoconus (KC), remains limited. The study aims to develop a new AI-based index (Keratoconus AI Index - KAI) to improve KC detection based on the tomographic data obtained by MS-39 (Costruzione Strumenti Oftalmici, CSO, Scandicci, Firenze, Italy).

This multicenter retrospective study was conducted in Humanitas Clinical and Research Center, Milan, Italy, Department of Ophthalmology, University Hospital Carl Gustav Carus, TU Dresden, Germany, Vissum Miranza, Alicante, Spain, Elza Institute, Zurich, Switzerland and Muscat Eye Laser Center, Muscat, Oman. 

A multilayer perceptron neural network with a backpropagation learning algorithm was trained on 80% of a dataset and validated on the remaining 20%. The dataset included 1,983 KC cases, 800 normal corneas, and 411 cases of very asymmetric ectasia (VAE). Diagnoses were confirmed by two masked corneal experts, followed by KISA analysis to ensure KISA% <60. A subsequent external validation was perfomed at Dresden University comparing the performance of KAI index vs the OCULUS TBI v2, BAD-D and CBI.

In the validation database, for keratoconus detection, KAI achieved 100% sensitivity and 97.45% specificity (cut-off 0.5). In distinguishing VAE from normal corneas, sensitivity was 84% and specificity 95.5% (cut-off 0.37), with an AUC of 0.968. The AI model provided robust separation between KC, VAE, and normal eyes, minimizing overlap. 

The KAI index demonstrates high accuracy in detecting KC and VAE, outperforming existing algorithms. Further studies are required to validate its clinical applicability.