Background and aims: Corneal ectasia risk assessment among elective refractive surgery candidateshas evolved to describe the cornea’s inherent susceptibility to biomechanical compensationand ectasia progression. The present study was designed to optimize artificial intelligence(AI) algorithms to integrate
corneal tomography and biomechanical data extracted based on theScheimpflug imaging technique and enhance the diagnosis of ectasia.Method: This cross-sectional, multicenter, case-control retrospective study included corneal datafrom ۳,۸۸۶ eyes of ۳,۴۱۲ patients without a history of corneal surgery. Corneal assessment wasperformed using Pentacam and Corvis ST (Oculus Optikgeräte GmbH; Wetzlar, Germany). Oneeye was randomly chosen from ۱,۶۸۰ normal subjects (N) and ۱,۱۸۱ patients with bilateral keratoconus(KC). In addition, ۵۵۱ eyes with normal placido-disk based corneal topography frompatients with very asymmetric ectasia (VAE-NT), and their ۴۷۴ ectatic eyes without prior cornealinterventions (VAE-E) were included. The current TBIv۱ (tomographic biomechanical index)was tested and an optimized AI algorithm was developed to increase the diagnostic accuracy ofectasia.Results: The diagnostic ability of the TBIv۱ expressed by the area under the receiver operatingcharacteristic curve (AUC) to differentiate clinical ectasia (KC and VAE-E) and VAE-NTwas ۰.۹۹۹ (with a sensitivity of ۹۸.۵%; ۹۸.۶% specificity for a cut-off value of ۰.۵) and ۰.۸۹۹(with a sensitivity of ۷۶%; ۸۹.۱% specificity for a cut-off value of ۰.۲۹), respectively. A novelrandom forest algorithm (TBIv۲ or BrAIN-TBI), developed with ۱۸ features in ۱۵۶ trees using۱۰-fold cross-validation, had a significantly higher AUC compared to TBIv۱ (۰.۹۴۵ versus ۰.۸۹۹,p<۰.۰۰۰۱) for detecting VAE-NT (with a sensitivity & specificity of ۸۴.۴% & ۹۰.۱%, for a cut-offvalue of ۰.۴۳), and similar AUC for clinical ectasia (۰.۹۹۹, p= ۰.۸۱۸; with a sensitivity & specificityof ۹۸.۷% & ۹۹.۲%, for a cut-off point of ۰.۸). TBIv۲ had a statistically higher AUC thanTBIv۱ when all cases were considered (۰.۹۸۵ versus ۰.۹۷۴, p< ۰.۰۰۰۱).Conclusion: Accuracy for ectasia detection is improved by AI integration of Scheimpflug-basedcorneal tomography and biomechanical assessments, which characterize ectasia susceptibility inthe varied VAE-NT group. True unilateral ectasia may occur in some VAE patients. Accuracy willbe continuously improved by machine learning when taking additional information into account,such as epithelial thickness or other indices from multimodal refractive imaging.