1Hothorn T., 1Lausen B., 1Adler W., 2Paulus D., 3Michelson G.

1Institut für Medizininformatik, Biometrie und Epidemiologie; 2Lehrstuhl für Mustererkennung, Institut für Informatik; 3Augenklinik mit Poliklinik, Universität Erlangen-Nürnberg, D-91054 Erlangen

Objective: To classify subjects as normal or glaucomatous based on laser scanning data from Heidelberg Retina Tomograph (HRT) and anamnestic data using tree classifiers.
Methods: In a cross-sectional study including 94 normal and 94 glaucoma subjects from the Erlangen Glaucoma Register (matched by age and sex) we evaluate the performance of tree based and linear discriminant analysis. The tree based methods are stabilized by bagging. Furthermore, the procedure of Swindale et al. (2000) is evaluted on our dataset. The error rates are estimated using 10% crossvalidation.
Results: Linear discriminant analysis classifies 22.22% of the eyes incorrect whereas pruned CART has an error rate of 20.44% and P-value adjusted classification trees give 18.28%. Bagging (Breiman, 1996) is able to reduce the error rate for pruned CART to 14.22%. Conclusion: The use of stabilized tree based methods decreases error rates for the classification of glaucoma based on HRT- and anamnestic data.
References: Breiman, L. (1996): Bagging Predictors, Machine Learning 26, 123-140. Swindale, N.V., Stjepanovic, G., Chin, A. and Mikelberg, F.S. (2000): Automated analysis of normal and glaucomatous optic nerve head topography images, IOVS 41(7), 1730-1742.

Copyright © Deutsche Ophthalmologische Gesellschaft, 2001. All rights reserved.

Last updated: 

Webdesign: SPALLEK.COM