Listy Biometryczne - Biometrical Letters, Vol. 40 (2003), No. 1, 15-22
|
ON RESULTS OBTAINED BY DISCRIMINANT ANALYSIS AND CLASSIFICATION TREES FOR MEDICAL DATA WITH CONSIDERABLY DIFFERENT GROUP SIZES
Piotr Jurkowski1, Malgorzata Ćwiklinska-Jurkowska2, Piotr Korbal3 1Laboratory of Informatics and Research Methodology, 2Department of Medical Informatics, 3Clinic of Obstetrics and Gynecology, The Ludwik Rydygier Medical University, ul. Skłodowskiej 9, 85-094 Bydgoszcz, Poland 1jurkomal@mail.atr.bydgoszcz.pl |
The paper investigates the results of discriminant analysis and classification trees, when sizes of groups are considerably different, on the basis of a real medical dataset. Probabilities a priori proportional to sizes of groups and cross-validation assessment of classification errors for Bayesian parametric and nonparametric discrimination and for classification trees are used. For considered methods very high specificity but low sensitivity was obtained. To be useful for aiding diagnosis, the classification procedure on the basis of the given database should be able to provide not only the satisfactory global classification error, but also good sensitivity and (or) specificity, depending on the character of medical problem or doctor's preferences. To achieve this goal one can try different actions and then check the chosen procedure by a testing sample.
discriminant analysis, classification trees, cross-validation, specificity, sensitivity, medical protection, low-birth weight.
