Uso de algoritmos de aprendizagem de máquina e estratégias de seleção de atributos para otimizar a identificação de ceracotone a partir de propriedades biomecânicas da córnea

Abstract

The purpose of the present study was to evaluate the Ocular Response Analyzer’s (ORA’s) performance in differentiating grades I and II keratoconus from normal corneas using each of its 41 parameters individually, and to assess the effect of analyzing all parameters together and the influence of the corneal thinnest point (CTP). In addition, investigate the use of machine learning algorithms and feature selection search strategies to optimize mild keratoconus identification. This study included 68 eyes with mild keratoconus (grades I and II) and 136 healthy agematched control eyes. All eyes had a central corneal thickness between 500 and 600 μm. The mean value of the 41 ORA parameters were compared between the groups, and between the subgroups created based on a CTP greater or lower than 500 μm. The area under the receiver operating characteristic curve (AUC) when separating both groups was calculated for each of the 41 parameters independently and for all of the parameters together. The 41 parameters were assessed using machine learning algorithms [support vector machine (SVM), decision tree, radial basis function neural network, and multi-layer perceptron] and feature selection strategies (forward selection, backward elimination, and genetic algorithm search). The algorithms’ performance was expressed as the sensitivity, specificity and accuracy obtained on the 10-fold cross-validation tests. Most parameters had a statistically lower mean value in the keratoconus group, although there was a large measurement overlap between both groups. Twenty-two parameters were not influenced by the CTP. When analyzed individually, 4 parameters (p1area, p1area1, p2area, and p2area1) had an AUC greater than 0.900. The p2area was the parameter that achieved the largest AUC individually (0.931). The AUC increased to 0.978 when analyzing all parameters together. Of the machine learning algorithms, SVM achieved the best performance when no feature selection strategies were used: sensitivity of 86.8%, specificity of 91.9%, and accuracy of 90.3% ± 5.2%. All algorithms had a better performance in mild keratoconus detection with the use of feature selection strategies. The highest sensitivity, specificity, and accuracy attained were of 94.1%, 95.6%, and 95.1% ± 4.5%, respectively. This performance was achieved with a subset of 23 ORA parameters selected by associating SVM with the genetic algorithm. This subset of parameters consisted of CRF, CH, aindex, p1area, p2area, aspect1, uslope2, dslope1, dslope2, w1, h2, mslew1, mslew2, slew1, slew2, p1area1, p2area1, uslope11, uslope21, dslope21, w11, path11 and path21. In conclusion, four ORA parameters were the best for identifying grades I and II keratoconus when used individually. The combination of all parameters improved the exam’s performance. The use of machine learning algorithms and feature selection search strategies optimized ORA’s detection of mild keratoconus.