Abstract
Purpose :
To compare the ability of discriminant functions and decision trees for building metrics to detect subclinical keratoconus.
Methods :
Group 1 consisted of 20 clinically normal eyes of 20 patients with newly diagnosed keratoconus in the fellow eye; for group 2 (control), 94 normal eyes were included. Anterior and posterior corneal surface Zernike coefficients (1st to 7th order, 6 mm analysis pupil) and pachymetry metrics were obtained by corneal Scheimpflug tomography (Pentacam HR). Single value metrics from input data were constructed with linear discriminant function analysis (SPSS 20) and decision trees (R statistical software package). Receiver operating characteristic curve (ROC) analysis was calculated for the output values of discriminant function analysis.
Results :
Accuracy for decision tree-based metrics ranged from 82.6% (anterior surface Zernike coefficients) to 90.4% (pachymetry), while metrics obtained with discriminant analysis reached an accuracy between 93.6% (anterior and posterior Zernike coefficients) to 98.9% (pachymetry and posterior surface Zernike coefficients). Specificity of decision trees was higher than the specificity of discriminant functions. Anterior Zernike coefficients achieved the highest specificity of 100 %. Sensitivity of decision trees was lower (0% to 75%) compared to sensitivity of discriminant function analysis (95% to 100%). Relevant Zernike coefficients were similar for both statistical methods (C3-1 anterior and posterior).
Conclusions :
Both decision trees and discriminant function analysis could distinguish between normal eyes and eyes with subclinical keratoconus. Although reaching high sensitivity, decision trees did not provide a clinically acceptable sensitivity.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.