Corneal Tomography Multivariate Index (CTMVI) effectively distinguishes healthy corneas from those susceptible to ectasia

• Autor(a) principal: de Almeida Jr, Gildasio Castello
• Data de Publicação: 2021
• Outros Autores: Guido, Rodrigo Capobianco [UNESP], Neto, Jogi Suda [UNESP], Rosa, João Marcos [UNESP], Castiglioni, Lilian, de Mattos, Luiz Carlos, Brandão, Cinara Cássia
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.bspc.2021.102995; http://hdl.handle.net/11449/233314
• topic: Artificial intelligence; Corneal ectasia; Corneal topography; Keratoconus; Paraconsistent feature engineering; Support vector machine

Background: Keratoconus (KC) is a multifactorial progressive corneal disease, pathogenically characterized by genetic, biochemical, and environmental key factors. Nevertheless, current tools used for its diagnosis, such as Pentacam, examine the posterior and anterior surfaces of the cornea based on a rotating Scheimpflug camera and provide, as output, a set of features for which the clinical interpretation has been error-prone. Objective: Consequently, the objective of this study is to describe a novel Artificial Intelligence (AI) strategy that receives subsets of the 52 original Pentacam tomographic descriptors as input and produces, as output, the scalar value called Corneal Tomography Multivariate Index (CTMVI). Specifically based on Paraconsistent Feature Engineering (PFE) and Support Vector Machine (SVM), it intends to help clinicians in the pre-diagnosis of cornea ectasia risk assessment. Methods: We collected data from patients’ eyes in three different classes: 411 with healthy corneas to compose the control group (CG), 302 with KC to represent the KC group (KCG) and, additionally, 64 in a group with very asymmetric ectasia but with normal corneal topography (VAE-NTG) in one eye, totaling 777 samples. A set of features from the patients’ corneal tomographic assessments was extracted to serve as input to the paraconsistent feature selector algorithm. Once it provides the best subsets of features representing the classes, a one-against-all SVM-based algorithm was successfully used to classify those groups. Notably, normal topography criteria (NTPC) were considered based on the steepest front curvature (Kmax) <47.2 diopters, a paracentral inferior–superior (I–S value) <1.45, and a keratoconus percentage index (KISA%) score <60. Normal corneal tomography (NTMC) consists of an anterior chamber depth <3.8μm, a front corneal apical elevation <4μm, a front corneal elevation (FCE) at the thinnest point <3.8μm, and an FCE <12μm in the corneal 4.0 mm center. Respective back corneal elevation values were 7μm, 13μm, and 25μm. To assess the clinical validity of the models and their abilities to correctly classify new data, two procedures were used: the first was a holdout cross validation, and the second was the external validation consisting of an independent test with samples that were not part of the training data. Results: Notably, PFE-based analysis effectively found dominant representative features which allowed for relevant classification accuracies based on SVM. Among numerous results, a parallel test combining CTMVI and existing Pentacam Random Forest Index (PRFI) produced a value of sensitivity of 0.99 and a value of specificity of 0.84, using 50% of the dataset for system modeling and the remaining 50% for assessment, improving current reports found in the literature. CTMVI, BAD-D, and PRFI showed an accuracy (acc) of 0.89, 0.86, and 0.89, respectively, in the external validation population to distinguish the healthy eyes from subclinical cases. The three indexes resulted in 0.99 acc when used to distinguish healthy eyes from corneal ectasia ones. CTMVI showed 100% sensitivity and 100% specificity between CG and KCG (cutoff ⩽−1.36). Regarding VAE-NTG patients, CTMVI had an area under the curve (AUC) of 0.935 (87.5% sensitivity; 84.95% specificity; cutoff ⩽0.639), which was higher than those of Belin/Ambrósio Enhanced Ectasia Total Deviation Value (BAD D) with AUC 0.765; 78.12% sensitivity; 65.05% specificity; p = .0001, and PRFI with AUC 0.805; 59.38% sensitivity; 90.78% specificity; p = .0004. During the external validation, it was shown to be as effective as BAD and PRFI. Conclusions: Based on the innovative joint application of PFE and SVM to distinguish healthy corneas from those pathologically-affected, we provided an original contribution to the state-of-the-art in terms of AI-aided diagnosis. Therefore, CTMVI improves the detection of corneal ectasia susceptibility using tomographic-derived descriptors. This reassures the efficacy of the proposed approach.