Modelos computacionais para otimização da escolha do anel intraestromal em pacientes com ceratocone utilizando dados tomográficos da córnea

Abstract

This work aims to improve the predictability of asphericity and average keratometry in keratoconus patients after implantation of intrastromal corneal ring segments (ICRS) by creating computational models based on machine learning, using tomographic data of the cornea. This study included 209 eyes of 160 keratoconus (grades I, II and III) implanted with ICRS. The Ferrara ICRS with 160 degrees of arch was implanted in all patients. The ICRS thickness varied from 150 to 250 micra. One or two segments were implanted. The base was composed of corneal tomography Pentacam® (Oculus, Wetzlar, Alemanha) parameters, clinical data and Ferrara ring data totaling 39 parameters. To create the models, neural network algorithms type multlayer perceptron (MLP) and linear regression were used. This study was conducted in four phases: (1) Preparation of the database and setting the values to be predicted mean keratometry and asphericity; (2) Calculation of the variation mean keratometry and asphericity and the nomogram calculation error; (3) Application of machine learning algorithms and attribute selection; (4) Mean keratometry and asphericity variation calculation provided for comparing algorithm with the variation of the preoperative and postoperative calculation of the algorithm and of the error. As a result, the best mean absolute error value found for asphericity was 0.19 and mean keratometry was 1.18. Comparing the mean absolute error values of the nomogram and the average absolute error of the algorithm, there was an improvement of 0.11 to asphericity and 0.09 to mean keratometry in relation to the current nomogram, confirming that the use of computational models can achieve more accurate results may contribute to surgical decision in an attempt to improve the quality of vision of keratoconus patients.