Enhanced Tomographic Assessment to Detect Corneal Ectasia Based on Artificial Intelligence.

PURPOSE To improve the detection of corneal ectasia susceptibility using tomographic data. DESIGN Multicenter case-control study. METHODS Data from patients from 5 different clinics from South America, the United States, and Europe were evaluated. Artificial intelligence (AI) models were generated using Pentacam HR (Oculus, Wetzlar, Germany) parameters to discriminate the preoperative data of 3 groups: stable laser-assisted in situ keratomileusis (LASIK) cases (2980 patients with minimum follow-up of 7 years), ectasia susceptibility (71 eyes of 45 patients that developed post-LASIK ectasia [PLE]), and clinical keratoconus (KC; 182 patients). Model accuracy was independently tested in a different set of stable LASIK cases (298 patients with minimum follow-up of 4 years) and in 188 unoperated patients with very asymmetric ectasia (VAE); these patients presented normal topography (VAE-NT) in 1 eye and clinically diagnosed ectasia in the other (VAE-E). Accuracy was evaluated with ROC curves. RESULTS The random forest (RF) provided highest accuracy among AI models in this sample with 100% sensitivity for clinical ectasia (KC+VAE-E; cutoff 0.52), being named Pentacam Random Forest Index (PRFI). Considering all cases, the PRFI had an area under the curve (AUC) of 0.992 (94.2% sensitivity, 98.8% specificity; cutoff 0.216), being statistically higher than the Belin/Ambrósio deviation (BAD-D; AUC = 0.960, 87.3% sensitivity, 97.5% specificity; P = .006, DeLong's test). The optimized cutoff of 0.125 provided sensitivity of 85.2% for VAE-NT and 80% for PLE, with 96.6% specificity. CONCLUSION The PRFI enhances ectasia diagnosis. Further integrations with corneal biomechanical parameters and with the corneal impact from laser vision correction are needed for assessing ectasia risk.