A comparative study of linear and quadratic discriminant classifier techniques for variable selection: a case study in predicting the effectiveness of hydroxyurea treatment of sickle cell anemia

Abstract

Compares the performance of linear and quadratic discriminant classifiers (LDC and QDC) in performing variable selection for computer assisted decision making. As a case study we used a medical database containing the initial and intermediate values for 23 mostly physiological measurements (variables) of sickle cell anemia patients recorded before and during hydroxyurea (HU) treatment. The variables selected by LDC and QDC were used to train three predictive models, quadratic discriminant analysis, multi-prototype classifier, and multi-layer perceptron to predict patients' response level to HU. The aim of this research is to provide the groundwork for further enhancement of artificial neural networks (ANNs) and other pattern recognition techniques in assisting in patient treatment assessment.