Four discriminant models for detecting keratoconus pattern using Zernike coefficients of corneal aberrations

Purpose: We compared the ability of four discriminant models to detect keratoconus (KC) using Zernike coefficients of corneal aberrations. Methods: We studied 51 eyes with KC, 46 with KC suspect, 50 after laser in situ keratomileusis, and 65 normal eyes. Four statistical discriminant analyses - linear discriminant analysis, k-nearest neighbor algorithm, Mahalanobis distance method, and neural network method - were performed using Zernike coefficients of corneal aberrations obtained by a Placido-based topographer. The detection scheme was constructed using a training set of data from one half of the randomly selected study participants, and performance was evaluated by a validation set in the other half. Results: Performance of the four models was different when <12 explanatory variables were included. Performance using the 2nd- to 4th-order Zernike terms did not differ significantly among models; average accuracy was 79 %. Conclusions: Determining explanatory variables of Zernike expansion coefficients of the corneal topography in discriminant models may contribute to improving accuracy of KC detection over the discriminant model, as appropriate selection of explanatory variables gave similar results despite different discriminant models.