External Validation Of The Artificial Intelligence Indices For Enhanced Detection Of Corneal Ectasia

To evaluate the diagnostic ability (area under the curve, AUC) of the optimized tomographic and biomechanical Index (TBIv2), current tomographic and biomechanical Index (TBIv1), and the Pentacam random forest index (PRFI), and the Belin/Ambrósio total deviation value (BAD-D) for the detection of early and clinically evident ectasia.

Department of Ophthalmology, University Hospital Carl Gustav Carus, TU Dresden, Germany

This prospective study enrolled 1984 eyes from 992 subjects from the refractive and keratoconus clinic between 2017 and 2020. Only one eye per subject was included and divided into three subgroups using the following inclusion criteria: 1) Normal eyes (NE): KISA% index<60, BAD-D<1.6, I-S value<1.45 D and Kmax<47 D; 2) bilateral keratoconus (KC): clinical signs of corneal ectasia in both eyes; 3) Very asymmetric ectasia (VAE-NT): the topographical normal eye from patients with clinical ectasia in the contralateral eye. Exclusion criteria were the presence of any ocular pathology except from KC, previous corneal or ocular surgeries, low quality of measurements, and diabetes mellitus. DeLong test was used to compare AUC of ROC analysis.

The final analysis included 177 NE, 403 KC, and 90 VAE-NT eyes. The TBIv2 had highest AUC of 0.987 in separating NE from all subgroups followed by PRFI (0.977), TBIv1 (0.969), and BAD-D (0.969). The AUC of TBIv2 was statistically significantly higher in comparison to all other parameters (all P<0.05). Separating NE from VAE-NT eyes, the TBIv2 statistically significantly outperformed all other parameters showing an AUC of 0.918 compared to 0.856, 0.809, and 0.803 for PRFI, TBIv1, and BAD-D, respectively (all P<0.05). A cut-off value of 0.58 (sensitivity 91%/specificity 100%) and 0.28 (87%/88%) was found for TBIv2 to separate NE and all subgroups as well as VAE-NT, respectively.

The TBIv2 had the best diagnostic ability to detect corneal ectasia. Further, the AUC of TBIv2 was higher than 0.9 in differentiating NE and VAE-NT, which could not be reached by using other parameters. The optimized TBI enhances the detection of subclinical cases. Some of the very asymmetric ectasia may be considered as unilateral ectasia. These findings support the concept that the optimized artificial intelligence for the integration of tomographical and biomechanical data enhances accuracy for screening ectasia risk prior laser vision.