New Glaucoma Classification Method based on Standard Heidelberg Retina Tomograph Parameters by Bagging Classification Trees

Purpose: In this article we propose and evaluate nonparametric tree classifiers that can handle non-normal data and a large number of possible predictors using the full set of standard Heidelberg Retina Tomograph measurements for classifying glaucoma. Methods: The classifiers were trained and tested using standard Heidelberg Retina Tomograph parameters from examinations of 98 subjects with glaucoma and 98 normal subjects of the Erlangen Glaucoma Registry. All patients and control subjects were evaluated by 15�-optic disc stereographs, Heidelberg Retina Tomograph measurements, standard computerized white-in-white perimetry, and 24-hour-intraocular pressure profiles. The subjects were matched by age and sex. Standard classification trees as well as bagged classification trees were used. The classification outcome of the trees was compared with the classification by two published linear discriminant functions based on Heidelberg Retina Tomograph variables with respect to their cross-validated misclassification error. Results: The bagged classification tree had the lowest misclassification error estimate of 14.8% with a sensitivity of 81.6% at a specificity of 88.8%. The cross-validated error rates of the two linear discriminant function procedures were 20.4% (sensitivity 82.6%, specificity 76.7%) and 20.6% (sensitivity 81.4%, specificity 77.3%) for our set of observations. Bagged classification trees were able to reduce the misclassification error of glaucoma classification. Conclusions: Bagged classification trees promise to be a new and efficient approach for glaucoma classification using morphometric 2- and 3-dimensional data derived from the Heidelberg Retina Tomograph, taking into account all given variables. Early treatment of glaucoma requires methods for diagnosing glaucoma before visual field defects are measurable. Various authors have proposed methods for classifying glaucoma based on Heidelberg Retina Tomograph (HRT) examinations. 1-7 The most commonly used classifiers were linear discriminant functions. This method is based on the assumption of multivariate normality and fails if the number of variables is large, whereas the number of observations is limited. Therefore, most authors select a small number of variables, which are incorporated into a linear discriminant analysis. However, this set of variables may not lead to the best linear classifier for other studies as shown recently. 8 In this article we propose and evaluate nonparametric tree classifiers, which can handle non-normal data and a large number of possible predictors without pre-selecting variables but using the full set of standard HRT measurements.