Comparing classification techniques for predicting essential hypertension

Hypertension is a leading cause of heart disease and stroke. In this study, performance of classification techniques is compared in order to predict the risk of essential hypertension disease. A retrospective analysis was performed in 694 subjects (452 patients and 242 controls). We compared performances of three decision trees, four statistical algorithms, and two neural networks. Predictor variables were age, sex, family history of hypertension, smoking habits, lipoprotein (a), triglyceride, uric acid, total cholesterol, and body mass index (BMI). Classification techniques were grouped using hierarchical cluster analysis (HCA). The data points appeared to cluster in three groups. The first cluster included MLP and RBF. Furthermore CART which was more similar than other techniques linked this cluster. The second cluster included FDA/MARS (degreeZ1), LR and QUEST, but FDA/MARS (degreeZ1) and LR was more similar than QUEST. The third cluster included FDA/MARS (degreeZ2), CHAID and FDA, but FDA/MARS (degreeZ2) and CHAID was more similar than FDA. MLP and RBF which are one each of neural networks procedures, performed better than other techniques in predicting hypertension. QUEST had a lesser performance than other techniques. q 2005 Elsevier Ltd. All rights reserved.